BPG is committed to discovery and dissemination of knowledge
Home / Archive / Volume 7, Issue 1
  • This Article
Table of Contents
Peer-Review Report of This Article
CrossCheck and Google Search of This Article
Academic Rules and Norms of This Article
Supplementary Materials of This Article
Citation of this article
Corresponding Author of This Article
Research Domain of This Article
Article-Type of This Article
Open-Access Policy of This Article
Times Cited Counts in Google of This Article
Journal Information of This Article
Publication Name
Artificial Intelligence in Cancer
ISSN
2644-3228
Publisher of This Article
Baishideng Publishing Group Inc, 7041 Koll Center Parkway, Suite 160, Pleasanton, CA 94566, USA

Retrospective Cohort Study Open Access
Copyright: ©Author(s) 2026. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution-NonCommercial (CC BY-NC 4.0) license. No commercial re-use. See permissions. Published by Baishideng Publishing Group Inc.
Artif Intell Cancer. Sep 8, 2026; 7(1): 116460
Published online Sep 8, 2026. doi: 10.35713/aic.v7.i1.116460
Magnetic resonance imaging-based deep learning model for prediction of the neoadjuvant chemotherapy response and survival prognosis in adolescents with osteosarcoma
Yu-Han Yang, West China School of Medicine, Sichuan University, Chengdu 6100041, Sichuan Province, China
ORCID number: Yu-Han Yang (0000-0002-4405-5711).
Author contributions: Yang YH contributed to study conception and design, data acquisition, analysis and interpretation, draft and critical revise the manuscript.
Institutional review board statement: This retrospective study involving human participants was reviewed and approved by the Institutional Review Boards of the participating institutions. All procedures were conducted in accordance with the ethical standards of the institutional and/or national research committees and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.
Informed consent statement: Patients were not required to give informed consent to the study because the analysis used anonymous clinical data that were obtained after each patient agreed to treatment by written consent.
Conflict-of-interest statement: The author reports no relevant conflicts of interest for this article.
STROBE statement: The authors have read the STROBE Statement-checklist of items, and the manuscript was prepared and revised according to the STROBE Statement-checklist of items.
Data sharing statement: De-identified individual participant data that underlie the results reported in this article are available from the corresponding author upon reasonable request. Data sharing is subject to approval by the relevant institutional review boards and execution of a data-use agreement to ensure protection of patient privacy and compliance with applicable regulations. Due to institutional policies and patient privacy considerations, raw imaging data or any data containing potentially identifying information will not be publicly released.
Corresponding author: Yu-Han Yang, West China School of Medicine, Sichuan University, No. 17 People’s South Road, Chengdu 6100041, Sichuan Province, China. yyh_1023@163.com
Received: November 21, 2025
Revised: December 3, 2025
Accepted: February 2, 2026
Published online: September 8, 2026
Processing time: 294 Days and 18.8 Hours

Abstract
BACKGROUND

Osteosarcoma is the most common primary bone malignancy in children, adolescents and young adults, and treatment typically includes preoperative neoadjuvant chemotherapy (NAC) followed by surgery. However, up to 20% of patients show resistance to NAC, exposing them to unnecessary toxicity and delayed definitive treatment. Noninvasive imaging biomarkers derived from magnetic resonance imaging (MRI), combined with deep learning (DL) methods, have the potential to capture intratumoral heterogeneity and enable accurate preoperative identification of chemotherapy responders and prediction of long-term outcomes.

AIM

To evaluate the prediction performance of the MRI-based model using DL methods for identification of response to NAC, and to explore the prognostic value of DL-based models for prediction of long-term survival in adolescents and young adults with osteosarcoma.

METHODS

All 134 eligible patients were included retrospectively from January 2012 to December 2018, including 94 patients in the training cohort and 40 patients in the testing cohort. We extracted DL-based features via transfer learning methods, and adopted support vector machine for MRI-based models construction evaluated by the area under the receiver operating characteristics curve (AUC). The MRI-based model with the highest AUC value was used to generate the DL-based signature. An integrated prediction model for response to NAC was developed including clinical variables and the DL-based signature. Additionally, the prognostic values of clinical variables and the DL-based signature were measured associated with overall survival (OS) by Cox proportional hazard analysis to develop an integrated prognostic model.

RESULTS

The integrated prediction model represented great discrimination abilities in the training cohort considering AUC of 0.961, accuracy of 90.43%, sensitivity of 92.45%, and specificity of 87.80%, while indicating AUC of 0.816, accuracy of 70.00%, sensitivity of 54.55%, and specificity of 88.89% in the testing cohort. The significant association has also been indicated in the integrated prognostic model with OS considering great classification and discrimination abilities.

CONCLUSION

The integrated prediction model represented effective abilities in identification of response to NAC, and the integrated prognostic model showed underlying prognostic value for OS in adolescents and young adults with osteosarcoma.

Key Words: Osteosarcoma; Magnetic resonance imaging; Neoadjuvant chemotherapy; Deep learning; Convolutional neural network

Core Tip: We developed and externally validated prediction models for histopathological diagnosis of resectable bone sarcomas using radiological features derived from both deep learning and handcrafted radiomics approaches. To our knowledge, this is the first study using pre-trained convolutional neural networks via transfer learning method to identify osteosarcoma and chondrosarcoma and predict long-term survival outcomes based on T1 and T2 magnetic resonance imaging images.
INTRODUCTION

Osteosarcoma is the most common primary bone sarcomas with one incidence peak in children and adolescents that these patients experience challenges during treatment with high tendencies of progression and metastasis[1]. Nowadays, the therapeutic strategy combining preoperative neoadjuvant chemotherapy (NAC), surgery, and postoperative treatments has improved the overall survivals (OS) for non-advanced osteosarcomas[2,3]. However, not all patients could benefit from this strategy that about 20% of patients were resistant to NAC inducing toxicities and delayed operation[4].

Accurate identification of pathological good responders (pGR) and non-pGR to NAC at early stage in adolescents and young adults with osteosarcoma could assist establishing appropriate treatments for improving survival rates and/or delaying disease progression[5].

Currently, the evaluation of tumor necrosis degree on postoperative histopathological analysis has been standards to assess responses to NAC[6]. The risk of injuries from needle or open biopsies and the errors from partial sampling might decrease the reliability of the current method. Therefore, a non-invasive tool was warranted to predict pathologic response prior to treatments. Non-invasive imaging examinations, especially magnetic resonance imaging (MRI) have been used to evaluate response to NAC in patients with osteosarcoma after at least one cycle of chemotherapy[7,8], which indicated that MRI imaging materials have enormous information for exploration and reflection of the intra-tumoral heterogeneity closely associated with NAC therapeutic response[8].

Deep learning (DL) methods could recognize images directly, represent accurate and consistent interpretations on imaging materials, and indicate associations with various clinical endpoints via multilayered convolutional neural networks (CNNs)[9,10]. DL-based radiomics methods have applied transfer learning on pre-trained CNNs to improve predictive abilities in small-sample datasets[11,12]. Combining feature extraction from pre-trained CNNs with traditional machine-learning classifiers can yield strong predictive performance while reducing computational burden for clinical applications[13]. The DL-based radiomics strategy has potential to predict response to NAC in adolescents and young adults with osteosarcoma.

The objectives of our study were to generate a noninvasive tool on preoperative MRI images to predict response to NAC in adolescents and young adults with osteosarcoma, and to combine clinical variables and the DL-based signature to construct integrated model in order to improve prediction capacities for NAC. We also aimed to construct DL-based models for predicting long-term prognosis to identify their prognostic value in adolescents and young adults with osteosarcoma receiving NAC.

MATERIALS AND METHODS
Patients

This retrospective study was approved by the Ethics Committee of the Institutional Review Board, and patients’ consent was waived because of the retrospective design. All eligible patients were diagnosed histologically as osteosarcomas with biopsy-proven evidence from January 2012 to December 2018, and received pre-treatment MRI examinations with available axial T1-weighted imaging (T1WI) and T2-weighted imaging with fat suppression (T2FS) within one month before treatments. All included patient received complete NAC before other treatments. Surgical resection was performed after NAC for evaluation of histopathological response to NAC. The exclusion criteria were as follow: (1) Receiving incomplete NAC or non-standard chemotherapy; (2) Without surgical specimens after NAC; (3) Unavailable pretreatment MRI images or missing MRI sequences; and (4) Low-quality MRI images (e.g., artifacts, low signal-to-noise ratio, low resolution and so on). Of the 134 patients included in the study, 94 (training cohort) were recruited from one institution and 40 (testing cohort) from another.

The clinical characteristics, including age, gender, tumor location, tumor size, and indicators from blood tests were collected from patients’ medical records. The MRI images acquisition protocols were shown in Supplementary Methods in Supplementary material. The NAC regimens were performed at least 4-6 cycles prior to surgical resection. The treatment protocol and timeline met the National Comprehensive Cancer Network guideline[14]. The NAC regimens were shown in Supplementary Methods in Supplementary material with details.

Clinical outcomes

The primary outcome was to identify the histopathological response to NAC based on surgical specimens after NAC. The pathologic assessment of response to NAC was conducted by two pathologists (one with 10-year experiences and one with 8-year experiences) based on the standard histological analysis from previous researches[6]. During histological analysis, the level of tumor necrosis exceeding 90% was considered as “pGR”, and the level of tumor necrosis less than 90% indicated “non-pGR”.

OS, the secondary outcome, was defined as the time from the date of histopathological diagnosis to the date of death, or last follow-up. The standard follow-up schedule was performed every 3 months for the first year, then every 6 months from the second year to the fifth year, and annually thereafter by telephone, mail, or out-patient. The last follow-up was censored in December, 2021.

Tumor segmentation

The regions of interest (ROIs) of primary tumor were delineated by two experienced radiologists (one with 10-year experiences and one with 15-year experiences) in MRI images separately using ITK-SNAP software[15]. Radiologists depicted the whole volumes of primary lesions for handcrafted radiomics features extraction. Thirty percent of patients in the training cohort were randomly and blindly selected to undergo repeat ROI segmentation one week after the initial segmentation. For assurance reproducibility for handcrafted radiomics features, the intra-class correlation coefficients were evaluated between inter-observer ROIs and between intra-observer ROIs. Only the handcrafted radiomics features with intra-class correlation coefficients larger than 0.85 were included for further analysis. The consecutive slices for the whole tumor volume were prepared for DL-based feature extraction. Each contoured slice was resized into 224 mm × 224 mm with a bounding box covering the automatic-delineated ROIs by DL algorithms ready as the input layer for pretrained CNNs. The detailed automatic segmentation procedure was listed in Supplementary Methods in Supplementary material.

DL-related features

We evaluated six widely used, pretrained CNN architectures: Xception[16], VGG16[17], VGG19[17], ResNet50[18], InceptionV3[19], and InceptionResNetV2[20] to extract DL features from axial T1WI and T2FS MRI sequences. All networks were initialized with weights from the ImageNet corpus to leverage transfer learning[21]. For each case, DL features were derived from ROI-centered inputs and subsequently subjected to feature-selection procedures and machine-learning model building (Supplementary Methods in Supplementary material for full details). Because the internal representations learned by these pretrained networks are complex and not directly interpretable, we applied Guided Gradient-weighted Class Activation Mapping to the final convolutional layer to localize image subregions that most strongly contributed to the extracted DL features[22].

Handcrafted radiomics

Handcrafted radiomics features were generated on radiologist-drawn ROIs using the PyRadiomics package (version 2.1.2) with and without wavelet filtration[23]. The study design complied with the image biomarker standardization initiative reporting guidelines that 851 features were extracted from each MRI sequence, including 107 features from original MRI images, and 744 features from wavelet filtered images. The handcrafted radiomic feature set comprised three categories: First-order statistical measures, morphological (shape) descriptors, and second-order texture features derived from matrix-based methods (gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size-zone matrix, gray-level dependence matrix and neighboring gray-tone difference matrix). Feature definitions and extraction procedures followed the image biomarker standardization initiative recommendations[24-26]; additional procedural information can be found in the Supplementary Methods in Supplementary material.

Prediction model for response to NAC

ComBat harmonization was applied to reduce variability introduced by MRI acquisition and reconstruction parameters and to ensure consistency of the derived imaging features[27]. We developed DL-based models to predict response to NAC and, for comparison, constructed models using handcrafted radiomics features. Feature selection in the training cohort proceeded in two stages: First, univariate analysis was used to rank features by area under the receiver operating characteristic curve (AUC), and the top 20% were retained; second, a wrapper approach based on recursive feature addition was employed to identify a final subset with high predictive AUC. Support vector machines (SVM) with a radial basis function kernel were used as the classification algorithm[28]. Model tuning and selection of optimal image pre-processing parameters were performed by nested cross-validation (ten repetitions of five-fold nested cross-validation) within the training set. Final models trained on the training cohort were evaluated on an independent external testing cohort. Further methodological details are provided in the Supplementary Methods in Supplementary material.

The deep-learning model that achieved the highest AUC in external validation was used to derive the DL signature. An integrated clinical-MRI nomogram combining this DL signature and selected clinical predictors was then developed using a SVM to maximize predictive performance. All clinical variables were first screened by univariate logistic regression (P < 0.10) and those meeting this threshold were entered into a multivariable logistic regression model. Only the clinical variables with P < 0.05 in the multivariate analysis were considered as clinical predictors. A clinical model including clinical predictors was constructed via the logistic regression method for comparison.

Prognostic model for OS

The DL-based signature from the optimal prediction model for response to NAC were applied to explore the prognostication for OS in adolescents and young adults with osteosarcoma. The DL-derived signature was dichotomized into low- and high-risk groups using an optimal cutoff of 0.90, determined from receiver operating characteristic (ROC) analysis of the training cohort with X-tile v3.6.1[29]. Clinical variables in the training cohort were first screened by univariate Cox proportional hazards analysis, with variables showing P < 0.10 entered into a multivariable model; variables with P < 0.05 in the multivariable analysis were retained as independent prognostic factors. An integrated Cox model combining these prognostic clinical variables and the DL signature was then constructed and presented as a nomogram to provide individualized OS estimates.

For assessment of prognostic performance for OS, Kaplan-Meier survival curves were generated and between-group differences - stratified by the DL-based signature and by pathological response to NAC - were tested using the log-rank test in both the training and testing cohorts. Calibration plots were used to compare predicted vs observed 3- and 5-year survival probabilities in each cohort. Discrimination and overall prognostic accuracy of the integrated Cox proportional hazards model and the DL-based signature were quantified using time-dependent concordance indices (C-index) and Brier scores at monthly intervals from 12 months to 90 months in the training and testing sets. Because the Brier score reflects both discrimination and calibration, smaller values indicate superior prognostic performance.

Statistical analysis

In both the training and testing cohorts, categorical variables were assessed with Fisher’s exact or χ2 tests, while continuous variables were analyzed using the nonparametric Mann-Whitney U test. The ROC curves and quantitative measurement indicators, including AUC value, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated to evaluate predictive performance of models for response to NAC. The calibration performance was measured by the Hosmer-Lemeshow test and visualized by calibration plots[30]. Discrimination abilities were assessed quantitatively by Harrell’s concordance indices (C-index)[31]. The clinical usefulness of the integrated nomogram model for response to NAC was evaluated by the net benefit of decision curve analysis[32]. Statistical significance was defined as a two-tailed P < 0.05. All statistical analyses and figure generation were performed using Python (v3.8) and R (v3.6.1). The specific software packages used are listed in the Supplementary Methods in Supplementary material.

RESULTS
Demographic characteristics

The overall workflow - comprising feature extraction, the model for predicting response to NAC, and the prognostic model for OS - is illustrated in Figure 1. This retrospective cohort included 134 patients, of whom 94 were allocated to the training set and 40 to the external testing set. Patient demographic and laboratory data are summarized in Table 1. Baseline clinical characteristics did not differ significantly between the training and testing cohorts (all P > 0.05), and the distribution of NAC response was comparable across cohorts (P = 0.883). There were 59 patients (44.0%) identified as pGR and 75 patients (56.0%) identified as non-pGR. In the training cohort, 41 patients were considered as pGR and 53 patients were considered as non-pGR, while 18 patients were considered as pGR and 22 patients were considered as non-pGR in the testing cohort. Patients’ age at diagnosis and tumor location were considered as clinical predictors in prediction of NAC in the training cohort (P < 0.05) (Supplementary Table 1).

Figure 1
Open in New Tab   Full Size Figure   Download Figure
Figure 1 A general flowchart of data analysis in prediction of response to neoadjuvant chemotherapy and survival outcome for adolescents and young adults with osteosarcoma. A: Features extraction: Imaging-derived features on T1- and T2- magnetic resonance imaging sequences were extracted by the deep learning (DL) analysis and handcrafted radiomics analysis, respectively; B: Prediction model construction and evaluation for response to neoadjuvant chemotherapy: The prediction models in identification of response to neoadjuvant chemotherapy were approached by machine learning methods via feature selection and model construction. The integrated nomogram model integrating clinical predictors and the DL-based signature was constructed to improve prediction performance. All prediction models were evaluated by prediction performance via receiver operating characteristic curves and calibration plots; C: The prognostic models construction and evaluation for overall survival: The DL-based signature from treatment prediction was dichotomized by the optimal cut-off points in association with overall survival with X-tile software. The integrated prognostic nomogram was constructed on prognostic clinical variables and the DL-based signature, and evaluated by Kaplan-Meier analysis, calibration curves, and time-dependent receiver operating characteristic analysis and Brier score. pGR: Pathological good response; SVM: Support vector machine; MRI: Magnetic resonance imaging; ROC: Receiver operating characteristic; DL: Deep learning; NAC: Neoadjuvant chemotherapy; AUC: Area under the receiver operating characteristics curve; KM: Kaplan-Meier.
Table 1 Demographic characteristics of adolescents and young adults with osteosarcoma in the training and testing cohort, n (%)/median (interquartile range).
Characteristic
All subjects (n = 134)
Training cohort (n = 94)
Testing cohort (n = 40)
P value
Response to neoadjuvant chemotherapy0.883
    pGR59 (44.0)41 (43.6)18 (45.0)
    Non-pGR75 (56.0)53 (56.4)22 (55.0)
Age, years16 (3)16 (3)17 (4)0.478
    ≤ 1497 (72.4)70 (74.5)27 (67.5)0.409
    > 1437 (27.6)24 (25.5)13 (32.5)
Gender0.184
    Female62 (46.3)47 (50.0)15 (37.5)
    Male72 (53.7)47 (50.0)25 (62.5)
Tumor size, cm7.3 (5.2)7.3 (4.7)8.1 (5.9)0.662
    ≤ 1098 (73.1)71 (75.5)27 (67.5)0.337
    > 1036 (26.9)23 (24.5)13 (32.5)
Location0.623
    Extremity78 (58.2)56 (59.6)22 (55.0)
    Trunk56 (41.8)38 (40.4)18 (45.0)
HGB1, g/L133 (26)133 (22)130 (28)0.399
    Abnormal41 (30.6)26 (27.7)15 (37.5)0.258
    Normal93 (69.4)68 (72.3)25 (62.5)
Platelet, 109/L211 (103)210 (102)219 (98)0.846
    Abnormal (> 300)26 (19.4)19 (20.2)7 (17.5)0.716
    Normal (≤ 300)108 (80.6)75 (79.8)33 (82.5)
WBC, 109/L5.84 (2.34)6.00 (2.30)5.68 (2.69)0.307
    Abnormal (≤ 4)20 (14.9)13 (13.8)7 (17.5)0.585
    Normal (> 4)114 (85.1)81 (86.2)33 (82.5)
ALB, g/L42.8 (6.2)42.8 (5.8)42.8 (7.0)0.967
    Abnormal (≤ 40)40 (29.9)26 (27.7)14 (35.0)0.395
    Normal (> 40)94 (70.1)68 (72.3)26 (65.0)
ALP, U/L99 (104)98 (101)114 (114)0.874
    Abnormal (> 140)42 (31.3)29 (30.9)13 (32.5)0.851
    Normal (≤ 140)92 (68.7)65 (69.1)27 (67.5)
LDH, U/L170 (77)169 (77)170 (77)0.948
    Abnormal (> 220)35 (26.1)25 (26.6)10 (25.0)0.847
    Normal (≤ 220)99 (73.9)69 (73.4)30 (75.0)
Duration of hospitalization, day15 (12)15 (12)15 (13)0.608
    ≤ 1466 (49.3)46 (48.9)20 (50.0)0.910
    > 1468 (50.7)48 (51.1)20 (50.0)
OS, month28 (24)30 (30)26 (17)0.192
1For male, hemoglobin < 130 is defined abnormal; for female, hemoglobin < 115 is defined abnormal.
pGR: Pathological good response; HGB: Hemoglobin; WBC: White blood cell; ALB: Serum albumin; ALP: Alkaline phosphatase; LDH: Lactate dehydrogenase; OS: Overall survival.
Open in New Tab Full Size Table
Prediction model for response to NAC

The clinical model, which incorporated age at diagnosis and tumor location, achieved an AUC of 0.647 and an accuracy of 47.87% in the training cohort, and an AUC of 0.729 with an accuracy of 52.50% in the testing cohort (Table 2). For the handcrafted radiomics models, four features derived from T1WI and one feature from T2FS were retained to build the combined radiomics model (Supplementary Table 2). This combined radiomics model yielded an AUC of 0.891 and an accuracy of 82.98% in the training cohort; in the testing cohort it produced an AUC of 0.588 (C-index: 0.588), accuracy: 55.00%, sensitivity: 50.00%, specificity: 61.11%, PPV: 61.11%, and NPV: 50.00% (Table 2).

Table 2 Predictive performance of radiological, clinical, and integrated nomogram models in prediction of response to neoadjuvant chemotherapy for adolescents and young adults with osteosarcoma in the training and testing cohorts.
ModelMRI sequence
Training cohort
Testing cohort
AUC
Accuracy
Sensitivity
Specificity
PPV
NPV
AUC
Accuracy
Sensitivity
Specificity
PPV
NPV
DL-based modelXceptionT10.98394.6896.2392.6894.4495.000.60462.5072.7350.0064.0060.00
T20.94488.3083.0295.1295.6581.250.50852.5022.7388.8971.4348.48
T1 + T20.98194.6892.4597.5698.0090.910.77077.5077.2777.7880.9573.68
VGG16T10.80475.5373.5878.0581.2569.570.56852.5054.5550.0057.1447.37
T20.88979.7973.5887.8088.6472.000.64962.5045.4583.3376.9255.56
T1 + T20.86779.7966.0497.5697.2268.970.54052.5036.3672.2261.5448.15
VGG19T10.88778.7269.8190.2490.2469.810.62657.5045.4572.2266.6752.00
T20.86278.7284.9170.7378.9578.380.54560.0086.3627.7859.3862.50
T1 + T20.88778.7269.8190.2490.2469.810.62657.5045.4572.2266.6752.00
ResNet50T10.90980.8583.0278.0583.0278.050.64960.0059.0961.1165.0055.00
T20.91281.9175.4790.2490.9174.000.63960.0063.6455.5663.6455.56
T1 + T20.94886.1783.0290.2491.6780.430.77065.0054.5577.7875.0058.33
InceptionV3T10.97790.4386.7995.1295.8384.780.68262.5077.2744.4462.9661.54
T20.96290.4392.4587.8090.7490.000.53560.0077.2738.8960.7158.33
T1 + T20.97889.3681.13100.00100.0080.390.56660.0068.1850.0062.5056.25
InceptionResNetV2T10.85077.6683.0270.7378.5776.320.74765.0086.3638.8963.3370.00
T20.82077.6686.7965.8576.6779.410.58167.5095.4533.3363.6485.71
T1 + T20.83076.6081.1370.7378.1874.360.71065.0086.3638.8963.3370.00
Handcrafted radiomics modelT10.897 80.8577.3685.3787.2374.470.63960.0045.4577.7871.4353.85
T20.903 80.8577.3685.3787.2374.470.66060.0040.9183.3375.0053.57
T1 + T20.891 82.9881.1385.3787.7677.780.58855.0050.0061.1161.1150.00
Clinical model1/0.647 47.8728.3073.1757.6944.120.72952.5036.3672.2261.5448.15
Nomogram model2/0.961 90.4392.4587.8090.7490.000.81670.0054.5588.8985.7161.54
1The clinical model was constructed by clinical variables selected by multivariate analysis with P value less than 0.05 in the training cohort.
2The nomogram model was constructed by significant clinical variables and the deep learning signature generated from the best deep learning magnetic resonance imaging (T1 + T2) model considering area under the receiver operating characteristic curve of the testing cohort among all models.
MRI: Magnetic resonance imaging; AUC: Area under the receiver operating characteristic curve; PPV: Positive predictive value; NPV: Negative predictive value; DL: Deep learning.
Open in New Tab Full Size Table

There were 18 DL-based models constructed for prediction of response to NAC, including six T1WI-derived models, six T2FS-derived models, and six combined models based on six pretrained CNNs. Among the DL-based models, the ResNet50-based combined model, which incorporated six MRI-derived features, achieved excellent performance in the training set (AUC: 0.948; accuracy: 86.17%; sensitivity: 83.02%; specificity: 90.24%; PPV: 91.67%; NPV: 80.43%). In external validation, this model yielded an AUC and C-index of 0.770, with accuracy 65.00%, sensitivity 54.55%, specificity 77.78%, PPV 75.00%, and NPV 58.33% (Table 2). The counts of DL-derived features retained for each model are listed in Supplementary Table 3. As for ResNet50 architecture, the performance of prediction models by features extracted from different ResNet50 layers verified the superiority of the current feature extraction strategy from the last layer before the fully connected layer with the highest AUCs in both training and testing cohorts among all ResNet50-based models (Supplementary Table 4). Guided Gradient-weighted Class Activation Mapping were generated to highlight important subregions for DL-based features (Figure 2). The specific feature patterns on feature maps were interpreted to valuable tumoral and peri-tumoral areas for further clinical view.

Figure 2
Open in New Tab   Full Size Figure   Download Figure
Figure 2 Feature heatmaps of representative patients on the deep learning ResNet50 algorithm via the Guided Gradient-weighted Class Activation Mapping in adolescents and young adults with osteosarcoma. A: Pathological good response (pGR); B: Non-pGR. The original magnetic resonance imaging images and their corresponding feature heatmaps were shown from left to right. The red color highlighted the region of interest to classify pGR and non-pGR to neoadjuvant chemotherapy. The red color focused on different area for pGR (A) and non-pGR (B) on T1-weighted magnetic resonance imaging sequence (left) and fat-saturated T2-weighted magnetic resonance imaging sequence (right) magnetic resonance imaging images, respectively. The concentration of these heatmaps preferred central area for non-pGR images, and lesion boundaries for pGR images. pGR: Pathological good response; T1WI: T1-weighted magnetic resonance imaging sequence; CAM: Class activation mapping; T2FS: Fat-saturated T2-weighted magnetic resonance imaging sequence.

The DL signature was derived from the ResNet50-based combined model. An integrated nomogram incorporating this DL signature and the clinical predictors age at diagnosis and tumor location was then developed (Figure 3A). In the training cohort the nomogram achieved an AUC of 0.961, accuracy 90.43%, sensitivity 92.45%, specificity 87.80%, PPV 90.74% and NPV 90.00%; in the independent testing cohort it attained an AUC (C-index) of 0.816, accuracy 70.00%, sensitivity 54.55%, specificity 88.89%, PPV 85.71% and NPV 61.54% (Table 2). The integrated model represented great classification performance by ROC curves (Figure 3B), and showed consistent calibration capacities (Figure 3C) and great clinical benefit (Figure 3D).

Figure 3
Open in New Tab   Full Size Figure   Download Figure
Figure 3 Evaluation of predictive performances for response to neoadjuvant chemotherapy for the integrated nomogram model in adolescents and young adults with osteosarcoma. A: Nomogram model combining significant clinical variables, age at diagnosis and tumor location, and the deep learning-based signature generated from the best deep learning-magnetic resonance imaging model considering area under the receiver operating characteristics curve of the testing cohort among all models; B: Receiver operating characteristic curves for the predictive performance of integrated nomogram model for response to neoadjuvant chemotherapy in the training and testing cohorts, respectively; C: Curves of the calibration analysis for the integrated nomogram model for response to neoadjuvant chemotherapy in the training and testing cohorts, respectively; D: The decision curve analysis of the integrated nomogram model for response to neoadjuvant chemotherapy. DL: Deep learning; ROC: Receiver operating characteristic.
Prognostic model for OS

The response to NAC was identified as prognostic clinical variables independently (Supplementary Table 5). The patients with non-pGR indicated higher risk of death compared patients with pGR [hazard ratio: 8.887, 95% confidence interval (CI): 1.123-70.325, P = 0.038]. An integrated prognostic nomogram was constructed by response to NAC and the DL-based signature via Cox proportional hazards in estimation of 3-year and 5-year OS (Figure 4A). The 3-year and 5-year survival probabilities for OS showed great calibration capacities (Figure 4B). To evaluate the discrimination performance of the integrated nomogram, the time-dependent AUC were evaluated (Figure 4C) with 3-year C-index of 0.773 (95%CI: 0.638-0.908) and 0.688 (95%CI: 0.421-0.955) in the training and testing cohorts, respectively. The 5-year C-index indicated 0.855 (95%CI: 0.725-0.984) vs 0.831 (95%CI: 0.527-1.000) in the training and testing cohorts, respectively. The time-dependent Brier scores were evaluated (Figure 4D) with 3-year Brier score of 14.9 (95%CI: 9.3-20.5) vs 16.1 (95%CI: 2.9-29.4), and 5-year Brier score of 16.6 (95%CI: 9.3-23.9) vs 20.8 (95%CI: 5.7-35.8) in the training and testing cohorts (Table 3). The prognostic value of the DL-based signature was evaluated slightly lower than the integrated nomogram model considering Brier score with great calibration performance (Table 3, Figure 4B). According to the results of Kaplan-Meier analysis, patients with lower DL-based signature yielded better OS than those with higher DL-based signature in both training (P < 0.001) and testing (P = 0.005) cohorts. The results also indicated the OS represented significant difference between pGR and non-pGR in the training cohort (89.136 vs 55.945, P < 0.001), not in the testing cohort (58.509 vs 63.341, P = 0.644) (Supplementary Figure 1, Supplementary Table 6).

Figure 4
Open in New Tab   Full Size Figure   Download Figure
Figure 4 Development and performance evaluation of the integrated prognostic model and the deep learning-based signature for prediction of overall survival in adolescents and young adults with osteosarcoma. A: Nomogram model in prediction of 3-year and 5-year overall survival (OS) combining prognostic clinical variables and the deep learning (DL)-based signature from treatment prediction in the training and testing cohort; B: Calibration curves in measurement of predicted 3-year (left) and 5-year (right) survival probabilities in the training (upper) and testing (lower) cohorts. The X-axis represented predicted survival risks, while the Y-axis showed observational survival probabilities. The line Y = X performed the ideal agreement between the estimated and actual survival probabilities; C: Time-dependent receiver operating characteristic analysis for the DL-based signature and the integrated nomogram model on OS showing the fluctuation of area under the receiver operating characteristics curves with follow-up in the training and testing cohorts; D: Time-dependent Brier scores for the DL-based signature and the integrated nomogram model on OS in the training and testing cohorts. NAC: Neoadjuvant chemotherapy; DL: Deep learning; OS: Overall survival; ROC: Receiver operating characteristic; AUC: Area under the receiver operating characteristic curve.
Table 3 Performance of the deep learning-based signature and integrated nomogram model in prediction of overall survival for adolescents and young adults with osteosarcoma in the training and testing cohorts.
ModelTimeTraining cohort
Testing cohort
C-index (95%CI)
Brier score (95%CI)
C-index (95%CI)
Brier score (95%CI)
Nomogram model13-year0.773 (0.638-0.908)14.9 (9.3-20.5)0.688 (0.421-0.955)16.1 (2.9-29.4)
5-year0.855 (0.725-0.984)16.6 (9.3-23.9)0.831 (0.527-1.000)20.8 (5.7-35.8)
DL-based signature23-year0.773 (0.638-0.908)14.9 (9.3-20.5)0.688 (0.421-0.955)17.0 (3.2-30.9)
5-year0.855 (0.725-0.984)16.8 (9.4-24.3)0.831 (0.527-1.000)21.9 (6.4-37.3)
1The nomogram model was constructed by prognostic clinical variables selected by multivariate Cox regression analysis on overall survival and the deep learning signature originated from diagnosis prediction.
2The deep learning signature was based on the best deep learning-radiological model considering AUC of the testing cohort from histological diagnosis identification.
C-index: Harrell’s concordance index; CI: Confidence interval; DL: Deep learning.
Open in New Tab Full Size Table
DISCUSSION

In this multicenter study, we developed prediction models for identification of response to NAC in adolescents and young adults with osteosarcoma on the MRI-based analysis, and validated the discriminative capacities in an independent cohort externally. The present study applied DL-based transfer learning technique to quantify and extract enormous DL-based features for interpretation of underlying tumor characterization. The integrated nomogram model on clinical predictors and the DL-based was present with improving performance in prediction of response to NAC superiority to the clinical predictors or the DL-based signature alone. In addition, an integrated nomogram for prognostication of OS combining prognostic clinical variables and the DL-based signature represented great prediction performance in both classification and discrimination after validation. The DL-based models on MRI images might serve as noninvasive tools for predicting treatment response and long-term prognosis in adolescents and young adults with osteosarcoma in further clinical practices.

Imaging modalities in routine examinations have been demonstrated to indicate treatment responses in osteosarcoma after development of radiomics techniques. The previous studies for prediction of chemotherapy showed favorable prediction performance in classification of responders and non-responders using contrast-enhanced computed tomography scans[33] and different MRI sequences[34-36]. Since MRI is an indispensable imaging procedure for diagnosis with high resolution on musculoskeletal tissues[14,37], MRI-based radiomics models have been applied in most cases to include more eligible patients and reach higher clinical validity compared with radiomics models on optional imaging modalities. However, it remains hard to identify patients with different responses to NAC accurately with small sample size due to the rarity of osteosarcoma and the complex acquisition procedure, especially in adolescents and young adults. To achieve excellent prediction abilities for response to NAC, the previous literatures have applied contrast utilization MRI images, which might decrease sample size and increase difficulties of imaging acquisition for adolescents and young adults[34-36]. The present study built radiomics-based models based on T1WI and T2FS for extensive clinical practicability in patients who are not tolerant of time-consuming and/or contrast-enhanced protocols. Therefore, our noninvasive model could be appropriate for adolescents and young adults, who might be hard to obtain quantified MRI images than older adults.

Recent advances in DL have the potential to substantially improve the performance and utility of models applied to medical imaging. However, training CNNs from scratch on single-center datasets is prone to overfitting, especially given the typically small number of annotated medical images available for specific clinical tasks. Using pretrained CNNs as feature extractors via transfer learning offers a practical remedy: It enables the adaptation of representations learned from large, general image corpora to the medical domain, thereby enhancing model generalizability and facilitating reproducibility across cohorts[12,38-40]. In this study, we demonstrate the effectiveness of transfer learning for addressing this clinical problem, as reflected by strong predictive performance. Despite these benefits, the internal feature-generation processes of pretrained CNNs remain difficult to interpret directly. By inspecting the feature maps, we localized the subregions that drive feature generation, thereby highlighting regions that demonstrate the capacity of deep-learning methods to recognize tumor imaging patterns. Figure 2 suggests that marginal zones of the lesion may be especially useful for detecting spatial heterogeneity within the tumor and its surrounding microenvironment.

In this study, we tried to combine clinical features and the DL-based signature for prediction of response to NAC to construct the integrated nomogram model for optimization of the prediction performance of the present MRI-based models. The prediction performance of our integrated nomogram model was reliable and cost effective, which might be an effective pretreatment measure for individualized therapy in adolescents and young adults with osteosarcoma. The integrated model assessed the intact lesions in order to decrease the uncertainty and increase the feasibility of treatment response, which could optimize pretreatment personalized therapeutic strategies to complement appropriate treatment for osteosarcoma.

There were some limitations existing in this study. First, despite using transfer learning, a mismatch remains between the domains of the pre-trained CNNs and our target dataset. A more definitive remedy would be to assemble large, task-specific image repositories with extensive expert annotations so that CNNs can be trained from scratch, improving both generalizability and clinical applicability. Second, our approach to ensuring feature robustness has limitations: Contour-based ROI methods may offer slightly lower repeatability than test-retest imaging protocols[24,41]. Implementing test-retest studies in a prospective setting would be more appropriate for validating feature stability, and our team plans to pursue this in future work.

In conclusion, this study constructed and validated DL-based models to accomplish pretreatment evaluation of response to NAC in adolescents and young adults with osteosarcoma. The combined clinical and deep-learning prediction model reliably classified pGR vs non-pGR after NAC, representing a new time-efficient method that does not require manual segmentation of tumor regions. The integrated clinical-DL prognostic model for OS was present with great prediction performance for long-term survival, which could assist therapy and surveillance in adolescents and young adults with osteosarcoma for improving patients’ outcomes.

CONCLUSION

In conclusion, this study constructed and validated DL-based models to accomplish pretreatment evaluation of response to NAC in adolescents and young adults with osteosarcoma. The integrated clinical-DL prediction model for response to NAC achieved the excellent performance in classification of pGR and non-pGR, which represented a novel manner with high efficiency and without manual tumor delineation. The integrated clinical-DL prognostic model for OS was present with great prediction performance for long-term survival, which could assist therapy and surveillance in adolescents and young adults with osteosarcoma for improving patients’ outcomes.

References
1.  Ritter J, Bielack SS. Osteosarcoma. Ann Oncol. 2010;21 Suppl 7:vii320-vii325.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 856]  [Cited by in RCA: 834]  [Article Influence: 52.1]  [Reference Citation Analysis (0)]
2.  Goorin AM, Schwartzentruber DJ, Devidas M, Gebhardt MC, Ayala AG, Harris MB, Helman LJ, Grier HE, Link MP; Pediatric Oncology Group. Presurgical chemotherapy compared with immediate surgery and adjuvant chemotherapy for nonmetastatic osteosarcoma: Pediatric Oncology Group Study POG-8651. J Clin Oncol. 2003;21:1574-1580.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 284]  [Cited by in RCA: 254]  [Article Influence: 11.0]  [Reference Citation Analysis (0)]
3.  Meyers PA, Schwartz CL, Krailo M, Kleinerman ES, Betcher D, Bernstein ML, Conrad E, Ferguson W, Gebhardt M, Goorin AM, Harris MB, Healey J, Huvos A, Link M, Montebello J, Nadel H, Nieder M, Sato J, Siegal G, Weiner M, Wells R, Wold L, Womer R, Grier H. Osteosarcoma: a randomized, prospective trial of the addition of ifosfamide and/or muramyl tripeptide to cisplatin, doxorubicin, and high-dose methotrexate. J Clin Oncol. 2005;23:2004-2011.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 517]  [Cited by in RCA: 530]  [Article Influence: 25.2]  [Reference Citation Analysis (0)]
4.  Bacci G, Ferrari S, Bertoni F, Ruggieri P, Picci P, Longhi A, Casadei R, Fabbri N, Forni C, Versari M, Campanacci M. Long-term outcome for patients with nonmetastatic osteosarcoma of the extremity treated at the istituto ortopedico rizzoli according to the istituto ortopedico rizzoli/osteosarcoma-2 protocol: an updated report. J Clin Oncol. 2000;18:4016-4027.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 298]  [Cited by in RCA: 277]  [Article Influence: 10.7]  [Reference Citation Analysis (0)]
5.  Luetke A, Meyers PA, Lewis I, Juergens H. Osteosarcoma treatment - where do we stand? A state of the art review. Cancer Treat Rev. 2014;40:523-532.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 1013]  [Cited by in RCA: 971]  [Article Influence: 80.9]  [Reference Citation Analysis (1)]
6.  Bacci G, Bertoni F, Longhi A, Ferrari S, Forni C, Biagini R, Bacchini P, Donati D, Manfrini M, Bernini G, Lari S. Neoadjuvant chemotherapy for high-grade central osteosarcoma of the extremity. Histologic response to preoperative chemotherapy correlates with histologic subtype of the tumor. Cancer. 2003;97:3068-3075.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 154]  [Cited by in RCA: 182]  [Article Influence: 7.9]  [Reference Citation Analysis (0)]
7.  Bajpai J, Gamnagatti S, Kumar R, Sreenivas V, Sharma MC, Khan SA, Rastogi S, Malhotra A, Safaya R, Bakhshi S. Role of MRI in osteosarcoma for evaluation and prediction of chemotherapy response: correlation with histological necrosis. Pediatr Radiol. 2011;41:441-450.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 81]  [Cited by in RCA: 87]  [Article Influence: 5.8]  [Reference Citation Analysis (0)]
8.  Byun BH, Kong CB, Lim I, Choi CW, Song WS, Cho WH, Jeon DG, Koh JS, Lee SY, Lim SM. Combination of 18F-FDG PET/CT and diffusion-weighted MR imaging as a predictor of histologic response to neoadjuvant chemotherapy: preliminary results in osteosarcoma. J Nucl Med. 2013;54:1053-1059.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 75]  [Cited by in RCA: 86]  [Article Influence: 6.6]  [Reference Citation Analysis (0)]
9.  LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436-444.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 70666]  [Cited by in RCA: 21437]  [Article Influence: 1948.8]  [Reference Citation Analysis (15)]
10.  Yamashita R, Nishio M, Do RKG, Togashi K. Convolutional neural networks: an overview and application in radiology. Insights Imaging. 2018;9:611-629.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 3779]  [Cited by in RCA: 1347]  [Article Influence: 168.4]  [Reference Citation Analysis (11)]
11.  Kermany DS, Goldbaum M, Cai W, Valentim CCS, Liang H, Baxter SL, McKeown A, Yang G, Wu X, Yan F, Dong J, Prasadha MK, Pei J, Ting MYL, Zhu J, Li C, Hewett S, Dong J, Ziyar I, Shi A, Zhang R, Zheng L, Hou R, Shi W, Fu X, Duan Y, Huu VAN, Wen C, Zhang ED, Zhang CL, Li O, Wang X, Singer MA, Sun X, Xu J, Tafreshi A, Lewis MA, Xia H, Zhang K. Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning. Cell. 2018;172:1122-1131.e9.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 3915]  [Cited by in RCA: 1868]  [Article Influence: 233.5]  [Reference Citation Analysis (5)]
12.  Shin HC, Roth HR, Gao M, Lu L, Xu Z, Nogues I, Yao J, Mollura D, Summers RM. Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning. IEEE Trans Med Imaging. 2016;35:1285-1298.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 4834]  [Cited by in RCA: 1979]  [Article Influence: 197.9]  [Reference Citation Analysis (4)]
13.  Raghu S, Sriraam N, Temel Y, Rao SV, Kubben PL. EEG based multi-class seizure type classification using convolutional neural network and transfer learning. Neural Netw. 2020;124:202-212.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 142]  [Cited by in RCA: 137]  [Article Influence: 22.8]  [Reference Citation Analysis (0)]
14.  Biermann JS, Chow W, Reed DR, Lucas D, Adkins DR, Agulnik M, Benjamin RS, Brigman B, Budd GT, Curry WT, Didwania A, Fabbri N, Hornicek FJ, Kuechle JB, Lindskog D, Mayerson J, McGarry SV, Million L, Morris CD, Movva S, O'Donnell RJ, Randall RL, Rose P, Santana VM, Satcher RL, Schwartz H, Siegel HJ, Thornton K, Villalobos V, Bergman MA, Scavone JL. NCCN Guidelines Insights: Bone Cancer, Version 2.2017. J Natl Compr Canc Netw. 2017;15:155-167.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 142]  [Cited by in RCA: 227]  [Article Influence: 25.2]  [Reference Citation Analysis (0)]
15.  Yushkevich PA, Piven J, Hazlett HC, Smith RG, Ho S, Gee JC, Gerig G. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage. 2006;31:1116-1128.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 5192]  [Reference Citation Analysis (0)]
16.  Chollet F.   Xception: Deep Learning with Depthwise Separable Convolutions. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017 Jul 21-26; Honolulu, HI, United States. Piscataway: IEEE, 2017: 1800-1807.  [PubMed]  [DOI]  [Full Text]
17.  Karen Simonyan AZ.   Very Deep Convolutional Networks for Large-Scale Image Recognition. International Conference on Learning Representations; 2015 May 7-9; San Diego, CA, United States. Washington: Computational and Biological Learning Society, 2025.  [PubMed]  [DOI]  [Full Text]
18.  Tank VH, Ghosh R, Gupta V, Sheth N, Gordon S, He W, Modica SF, Prestigiacomo CJ, Gandhi CD. Drug eluting stents versus bare metal stents for the treatment of extracranial vertebral artery disease: a meta-analysis. J Neurointerv Surg. 2016;8:770-774.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 52]  [Cited by in RCA: 45]  [Article Influence: 4.5]  [Reference Citation Analysis (0)]
19.  Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z.   Rethinking the Inception Architecture for Computer Vision. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2016; 2016 Jun 27-30; Las Vegas, NV, United States. Piscataway: IEEE, 2016.  [PubMed]  [DOI]  [Full Text]
20.  Szegedy C, Ioffe S, Vanhoucke V, Alemi A.   Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning. Proceedings of the AAAI Conference on Artificial Intelligence; 2017 Feb 4-9; San Francisco, CA, United States. Washington; AAAI Press, 2017.  [PubMed]  [DOI]  [Full Text]
21.  Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang ZH, Karpathy A, Khosla A, Bernstein M, Berg AC, Li FF.   ImageNet Large Scale Visual Recognition Challenge. International Conference on Computer Vision and Pattern Recognition (CVPR 2015); 2015 Jun 7-12; Boston, MA, United States. Piscataway: IEEE, 2015.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 16940]  [Cited by in RCA: 6594]  [Article Influence: 599.5]  [Reference Citation Analysis (0)]
22.  Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. Int J Comput Vis. 2020;128:336-359.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 977]  [Cited by in RCA: 1709]  [Article Influence: 244.1]  [Reference Citation Analysis (0)]
23.  van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts HJWL. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res. 2017;77:e104-e107.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 5416]  [Cited by in RCA: 4709]  [Article Influence: 523.2]  [Reference Citation Analysis (4)]
24.  Zwanenburg A, Leger S, Agolli L, Pilz K, Troost EGC, Richter C, Löck S. Assessing robustness of radiomic features by image perturbation. Sci Rep. 2019;9:614.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 97]  [Cited by in RCA: 196]  [Article Influence: 28.0]  [Reference Citation Analysis (3)]
25.  Fiset S, Welch ML, Weiss J, Pintilie M, Conway JL, Milosevic M, Fyles A, Traverso A, Jaffray D, Metser U, Xie J, Han K. Repeatability and reproducibility of MRI-based radiomic features in cervical cancer. Radiother Oncol. 2019;135:107-114.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 77]  [Cited by in RCA: 131]  [Article Influence: 18.7]  [Reference Citation Analysis (0)]
26.  Zwanenburg A, Vallières M, Abdalah MA, Aerts HJWL, Andrearczyk V, Apte A, Ashrafinia S, Bakas S, Beukinga RJ, Boellaard R, Bogowicz M, Boldrini L, Buvat I, Cook GJR, Davatzikos C, Depeursinge A, Desseroit MC, Dinapoli N, Dinh CV, Echegaray S, El Naqa I, Fedorov AY, Gatta R, Gillies RJ, Goh V, Götz M, Guckenberger M, Ha SM, Hatt M, Isensee F, Lambin P, Leger S, Leijenaar RTH, Lenkowicz J, Lippert F, Losnegård A, Maier-Hein KH, Morin O, Müller H, Napel S, Nioche C, Orlhac F, Pati S, Pfaehler EAG, Rahmim A, Rao AUK, Scherer J, Siddique MM, Sijtsema NM, Socarras Fernandez J, Spezi E, Steenbakkers RJHM, Tanadini-Lang S, Thorwarth D, Troost EGC, Upadhaya T, Valentini V, van Dijk LV, van Griethuysen J, van Velden FHP, Whybra P, Richter C, Löck S. The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping. Radiology. 2020;295:328-338.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 3143]  [Cited by in RCA: 3028]  [Article Influence: 504.7]  [Reference Citation Analysis (13)]
27.  Hu Y, Xie C, Yang H, Ho JWK, Wen J, Han L, Lam KO, Wong IYH, Law SYK, Chiu KWH, Vardhanabhuti V, Fu J. Computed tomography-based deep-learning prediction of neoadjuvant chemoradiotherapy treatment response in esophageal squamous cell carcinoma. Radiother Oncol. 2021;154:6-13.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 38]  [Cited by in RCA: 112]  [Article Influence: 22.4]  [Reference Citation Analysis (0)]
28.  Amari S, Wu S. Improving support vector machine classifiers by modifying kernel functions. Neural Netw. 1999;12:783-789.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 659]  [Cited by in RCA: 224]  [Article Influence: 8.3]  [Reference Citation Analysis (0)]
29.  Camp RL, Dolled-Filhart M, Rimm DL. X-tile: a new bio-informatics tool for biomarker assessment and outcome-based cut-point optimization. Clin Cancer Res. 2004;10:7252-7259.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 3094]  [Cited by in RCA: 3152]  [Article Influence: 143.3]  [Reference Citation Analysis (3)]
30.  Vickers AJ, Cronin AM, Elkin EB, Gonen M. Extensions to decision curve analysis, a novel method for evaluating diagnostic tests, prediction models and molecular markers. BMC Med Inform Decis Mak. 2008;8:53.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 1046]  [Cited by in RCA: 1074]  [Article Influence: 59.7]  [Reference Citation Analysis (4)]
31.  Schoop R, Beyersmann J, Schumacher M, Binder H. Quantifying the predictive accuracy of time-to-event models in the presence of competing risks. Biom J. 2011;53:88-112.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 41]  [Cited by in RCA: 54]  [Article Influence: 3.6]  [Reference Citation Analysis (0)]
32.  Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26:565-574.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 4224]  [Cited by in RCA: 4360]  [Article Influence: 218.0]  [Reference Citation Analysis (5)]
33.  Lin P, Yang PF, Chen S, Shao YY, Xu L, Wu Y, Teng W, Zhou XZ, Li BH, Luo C, Xu LM, Huang M, Niu TY, Ye ZM. A Delta-radiomics model for preoperative evaluation of Neoadjuvant chemotherapy response in high-grade osteosarcoma. Cancer Imaging. 2020;20:7.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 41]  [Cited by in RCA: 94]  [Article Influence: 15.7]  [Reference Citation Analysis (0)]
34.  Baidya Kayal E, Kandasamy D, Khare K, Bakhshi S, Sharma R, Mehndiratta A. Texture analysis for chemotherapy response evaluation in osteosarcoma using MR imaging. NMR Biomed. 2021;34:e4426.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 11]  [Cited by in RCA: 29]  [Article Influence: 5.8]  [Reference Citation Analysis (0)]
35.  Lee SK, Jee WH, Jung CK, Im SA, Chung NG, Chung YG. Prediction of Poor Responders to Neoadjuvant Chemotherapy in Patients with Osteosarcoma: Additive Value of Diffusion-Weighted MRI including Volumetric Analysis to Standard MRI at 3T. PLoS One. 2020;15:e0229983.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 10]  [Cited by in RCA: 23]  [Article Influence: 3.8]  [Reference Citation Analysis (0)]
36.  Chen H, Zhang X, Wang X, Quan X, Deng Y, Lu M, Wei Q, Ye Q, Zhou Q, Xiang Z, Liang C, Yang W, Zhao Y. MRI-based radiomics signature for pretreatment prediction of pathological response to neoadjuvant chemotherapy in osteosarcoma: a multicenter study. Eur Radiol. 2021;31:7913-7924.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 14]  [Cited by in RCA: 52]  [Article Influence: 10.4]  [Reference Citation Analysis (0)]
37.  Strauss SJ, Frezza AM, Abecassis N, Bajpai J, Bauer S, Biagini R, Bielack S, Blay JY, Bolle S, Bonvalot S, Boukovinas I, Bovee JVMG, Boye K, Brennan B, Brodowicz T, Buonadonna A, de Álava E, Dei Tos AP, Garcia Del Muro X, Dufresne A, Eriksson M, Fagioli F, Fedenko A, Ferraresi V, Ferrari A, Gaspar N, Gasperoni S, Gelderblom H, Gouin F, Grignani G, Gronchi A, Haas R, Hassan AB, Hecker-Nolting S, Hindi N, Hohenberger P, Joensuu H, Jones RL, Jungels C, Jutte P, Kager L, Kasper B, Kawai A, Kopeckova K, Krákorová DA, Le Cesne A, Le Grange F, Legius E, Leithner A, López Pousa A, Martin-Broto J, Merimsky O, Messiou C, Miah AB, Mir O, Montemurro M, Morland B, Morosi C, Palmerini E, Pantaleo MA, Piana R, Piperno-Neumann S, Reichardt P, Rutkowski P, Safwat AA, Sangalli C, Sbaraglia M, Scheipl S, Schöffski P, Sleijfer S, Strauss D, Sundby Hall K, Trama A, Unk M, van de Sande MAJ, van der Graaf WTA, van Houdt WJ, Frebourg T, Ladenstein R, Casali PG, Stacchiotti S; ESMO Guidelines Committee, EURACAN, GENTURIS and ERN PaedCan. Bone sarcomas: ESMO-EURACAN-GENTURIS-ERN PaedCan Clinical Practice Guideline for diagnosis, treatment and follow-up. Ann Oncol. 2021;32:1520-1536.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 123]  [Cited by in RCA: 350]  [Article Influence: 70.0]  [Reference Citation Analysis (9)]
38.  Lopes UK, Valiati JF. Pre-trained convolutional neural networks as feature extractors for tuberculosis detection. Comput Biol Med. 2017;89:135-143.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 131]  [Cited by in RCA: 107]  [Article Influence: 11.9]  [Reference Citation Analysis (0)]
39.  Yun J, Park JE, Lee H, Ham S, Kim N, Kim HS. Radiomic features and multilayer perceptron network classifier: a robust MRI classification strategy for distinguishing glioblastoma from primary central nervous system lymphoma. Sci Rep. 2019;9:5746.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 65]  [Cited by in RCA: 72]  [Article Influence: 10.3]  [Reference Citation Analysis (0)]
40.  Zhu Y, Man C, Gong L, Dong D, Yu X, Wang S, Fang M, Wang S, Fang X, Chen X, Tian J. A deep learning radiomics model for preoperative grading in meningioma. Eur J Radiol. 2019;116:128-134.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 63]  [Cited by in RCA: 124]  [Article Influence: 17.7]  [Reference Citation Analysis (0)]
41.  Balagurunathan Y, Kumar V, Gu Y, Kim J, Wang H, Liu Y, Goldgof DB, Hall LO, Korn R, Zhao B, Schwartz LH, Basu S, Eschrich S, Gatenby RA, Gillies RJ. Test-retest reproducibility analysis of lung CT image features. J Digit Imaging. 2014;27:805-823.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 163]  [Cited by in RCA: 196]  [Article Influence: 17.8]  [Reference Citation Analysis (0)]
Footnotes

Peer review: Externally peer reviewed.

Peer-review model: Single blind

Specialty type: Orthopedics

Country of origin: China

Peer-review report’s classification

Scientific quality: Grade B, Grade B

Novelty: Grade B, Grade B

Creativity or innovation: Grade B, Grade B

Scientific significance: Grade B, Grade B

P-Reviewer: Qu CS, Chief, China S-Editor: Hu XY L-Editor: A P-Editor: Xu J

Write to the Help Desk
Copyright © 1993-2026 Baishideng Publishing Group Inc, 7041 Koll Center Parkway, Suite 160, Pleasanton, CA 94566, USA. All rights reserved, including rights relating to text and data mining, AI training, and similar technologies. For open-access content, the applicable copyright and licensing terms govern.