Development and validation of a multimodal MRI habitat-based deep learning fusion model for predicting 252Cf neutron therapy response in cervical cancer

Abstract

BACKGROUND

Cervical cancer remains one of the leading causes of cancer-related mortality among women worldwide. Although californium-252 (²⁵²Cf) neutron brachytherapy has demonstrated favorable therapeutic efficacy for locally advanced cervical cancer, substantial inter-patient heterogeneity in treatment response still exists. Therefore, developing a reliable non-invasive predictive model before treatment is of considerable clinical importance.

AIM

To develop and validate a predictive model that integrates multiparametric magnetic resonance imaging (MRI) habitat atlas, radiomics, and deep learning features for the non-invasive and accurate prediction of treatment efficacy in cervical cancer patients receiving ²⁵²Cf neutron intracavitary brachytherapy combined with external beam radiotherapy.

METHODS

A total of 100 cervical cancer patients who underwent the aforementioned treatment at our institution from January 2020 to June 2025 were retrospectively enrolled. All patients underwent pretreatment multi-sequence MRI scans [T1-weighted imaging, T2-weighted imaging, dynamic contrast enhanced (DCE)-MRI, and diffusion weighted imaging (DWI)]. Based on DCE-MRI and DWI parameter maps, a k-means clustering algorithm was employed to delineate the tumor habitat atlas, partitioning the tumor into three functional subregions (H1: High vascularity/high cellularity; H2: Low vascularity/high cellularity; H3: Low vascularity/low cellularity). Radiomics features were extracted from both the whole tumor region and each habitat subregion, while deep learning features were extracted using a pre-trained ResNet50 network. The aforementioned features, along with clinical-dosimetric parameters, were fused. Following feature selection via the least absolute shrinkage and selection operator, a fusion prediction model was constructed using a support vector machine. The dataset was split into a training set (n = 70) and a validation set (n = 30) at a 7:3 ratio for model evaluation.

RESULTS

Habitat analysis successfully identified three types of functional subregions, whose distribution was significantly associated with tumor heterogeneity. The fusion model (Model-Fusion) achieved an area under curve (AUC) of 0.892 (95% confidence interval: 0.821-0.963) and an accuracy of 0.867 in the validation set, significantly outperforming the model based solely on clinical features (AUC = 0.712), the whole-tumor radiomics model (AUC = 0.783), and the radiomics combined with deep learning model (AUC = 0.835) (all P < 0.05). Decision curve analysis demonstrated that the fusion model provided the highest clinical net benefit. The 3-year overall survival rate was significantly higher in the model-predicted high-benefit group (92.5%) compared to the low-benefit group (71.4%, P = 0.003).

CONCLUSION

This study integrates MRI habitat atlas analysis with deep learning and radiomics. The constructed multimodal fusion model can non-invasively and accurately predict the efficacy of ²⁵²Cf neutron therapy for cervical cancer, providing a promising approach for individualized precision radiotherapy decision-making.

Key Words: Cervical cancer; Magnetic resonance imaging; Habitat analysis; Deep learning; Radiomics; Californium-252 neutron therapy; Treatment efficacy prediction

Core Tip: The present study developed a multimodal fusion predictive model integrating MRI habitat atlas, deep learning, and radiomics to fill the current gap in early and non-invasive prediction of treatment efficacy of californium-252 (²⁵²Cf) neutron therapy for cervical cancer, thereby provide decisive evidence and a powerful imaging tool for individualized precision radiotherapy decision-making.

INTRODUCTION

Cervical cancer is one of the most common malignant cancers affecting women’s health globally, with 600000 new cases reported worldwide in 2020. Up to now, concurrent chemoradiotherapy remains the standard treatment for cervical cancer. Californium-252 (²⁵²Cf) neutron intracavitary brachytherapy has been found to show great potential in extending long-term survival when treating patients with locally advanced cervical cancer owing to its high linear energy transfer (LET, about 85 keV/μm), high relative biological effectiveness (RBE, around 3-5), and distinctive advantages for killing hypoxic tumor cells[1]. Although high-LET neutron irradiation is less influenced by hypoxia than conventional photon radiotherapy, treatment response remains highly heterogeneous[2-4]. Increasing evidence suggests that this variability is associated with multiple biological factors, including intratumoral heterogeneity[5], enhanced DNA damage repair capacity[6], cancer stem cell populations[7], and remodeling of the tumor microenvironment[8]. All these emphasize the critical importance of pretreatment non-invasive and accurate prediction of treatment outcomes to identify patients who may benefit from the treatment, and to optimize personalized care.

Traditionally, treatment efficacy has been predicted mainly according to clinical indicators such as cancer staging, pathological types, and serum markers. This approach, however, treats tumors as uniform entities and ignores their inherent spatial heterogeneity, resulting in limited prediction accuracy. In contrast, multiparametric magnetic resonance imaging (MRI) can present plentiful vivid morphological and functional imaging information. Radiomics enables high-throughput extraction of quantitative image features, while deep learning allows for automatic mining of underlying abstract features. These two pretreatment MRI data-based frameworks have exhibited promising feasibility in predicting treatment response to chemoradiotherapy in cervical cancer[9-11].

Based on physiological parameters acquired using multiparametric MRI, the novel technique tumor habitat analysis has the ability to partition intra-tumoral environment into functional subregions (habitat atlas), which may have different biological characteristics, through unsupervised clustering methods, thus offering more nuanced characterization of tumor microenvironment heterogeneity[5,12]. Previous studies have demonstrated that habitat analysis is superior to the conventional whole tumor analysis in prediction of reoccurrence after radiotherapy for brain metastasis or response to neoadjuvant chemotherapy for breast cancer[13,14].

We hypothesized that integrating habitat-specific MRI information with radiomics and deep learning features would significantly improve prediction performance compared with conventional imaging models. Up to now, however, no study has applied this integrated approach to predict the efficacy of ²⁵²Cf neutron therapy for cervical cancer. Accordingly, the present study aimed to addresses this gap by constructing a multimodal fusion predictive model for early and accurate prediction of treatment efficacy of ²⁵²Cf neutron therapy for cervical cancer, with the goal of providing evidence to support personalized precision radiotherapy.

MATERIALS AND METHODS

Participants

Data were collected on patients who received ²⁵²Cf neutron brachytherapy for cervical cancer at our hospital between January 2020 to December 2025. The Ethics Committee of the Third Affiliated Hospital of Qiqihar Medical University has reviewed and approved this study (Approval No. 2025 LL-208). All study participants or their legal guardian provided informed written consent about personal and medical data collection prior to study enrolment.

The inclusion criteria were: (1) Cervical squamous cell carcinoma or cervical adenocarcinoma confirmed by histopathology; (2) Underwent multiparametric pelvic MRI at 3.0T within two weeks prior to treatment, with complete sequences including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), dynamic contrast enhanced (DCE)-MRI, and diffusion weighted imaging (DWI); (3) Stage IB to IVA according to the 2018 International Federation of Gynecology and Obstetrics (FIGO) staging system for cervical cancer; (4) Had complete dosimetric parameters for ²⁵²Cf neutron therapy; and (5) Underwent regular follow-up for at least 6 months after treatment, with comprehensive efficacy assessments.

The exclusion criteria were: (1) Poor image quality on MRI scans (motion artifact score > 2) or missing key sequences; (2) Previous pelvic radiotherapy or presence of other malignant tumors; (3) Incomplete clinical or follow-up data; and (4) Pregnancy.

In total, 100 patients were included in the study. The were assigned to a training set (70 patients) and a validation set (30 patients) at a ratio of 7:3 using the stratified randomization method. Randomization was stratified by FIGO stage and treatment response to ensure baseline comparability between the two groups.

Chemoradiotherapy

All patients received pelvic external beam radiation combined with ²⁵²Cf neutron brachytherapy and concurrent chemotherapy with cisplatin (40 mg/m², once a week). Pelvic external beam radiation was delivered using a 6-MV X-band linear accelerator was used with conventional fractionation delivered (1.8-2.0 Gy/fraction), totaling 46-50 Gy over 23-25 fractions. For ²⁵²Cf neutron brachytherapy, the ZH-1000 afterloading machine (Manufacturer, Zunrui Keji) was used, with an intrauterine tube combined with a vaginal ovoid applicator. Neutron brachytherapy was delivered at 8-10 Gy per insertion per week (single RBE-weighted dose) for 3-4 insertions, resulting in an overall RBE-weighted dose of 30-40 Gy to reference point A (RBE = 3.0). All brachytherapy procedures were performed based on 3D computed tomography/MRI-guided planning, with dose optimized via the Prague-Swedish traction power supply system to achieve the prescribed dose constraints, with a D90 for the high-risk clinical target volume of ≥ 85 Gy in EQD2[15,16].

Image acquisition and preprocessing

Patients were positioned supine and scanned on an empty stomach using Siemens MRI scanners. The pulse sequences and parameters were as follows: (1) Fat-saturated T2WI [repetition time (TR)/echo time (TE) = 4200/90 ms, slice thickness = 3 mm, inter-slice gap = 0 mm, field of view (FOV) = 24 × 24 cm, matrix = 384 × 384]; (2) Axial DWI with echo-planar imaging (EPI) (b-values = 0 and 800 s/mm², TR/TE = 4800/63 ms, slice thickness = 3 mm); (3) DCE-MRI (temporal resolution ≈ 6 s, 20 phases, TR/TE = 4.76/1.68 ms, with intravenous injection of Gd-DTPA at 2 mL/s); and (4) Axial T1WI (TR/TE = 560/10 ms).

Image preprocessing was performed as follows: (1) DICOM images were converted to NIfTI format using SimpleITK; (2) Multi-sequence rigid registration was performed using ANTs, with T2WI as the reference; (3) Two experienced radiologists (with 10 and 8 years of experience in pelvic MRI, respectively) manually delineated the whole-organ volume of interest (VOI) on T2WI images. The delineation was transferred via the radiotherapy planning system (Eclipse v16.0) using Seq2Seq mapping. Inter-observer agreement was assessed using the intraclass correlation coefficient (ICC > 0.85), and the final regions of interest were obtained by averaging the two delineations; (4) N4 bias field correction was applied to MRI images; and (5) All images were resampled to a uniform resolution (voxel size: 1 × 1 × 1 mm³) and normalized using z-standardization to generate z-scores[17,18].

Tumor habitat delineation

Tumor habitats were identified based on DCE-MRI pharmacokinetic parameters [volume transfer constant (Ktrans), extracellular volume fraction (Ve), derived from the Tofts two-compartment model] and DWI-derived parameters [apparent diffusion coefficient (ADC)], and were voxel-wise extracted within the tumor VOI.

Feature extraction

Radiomic features: A total of 1218 radiomic features, amounting to 4872 habitat dimensions (including habitat features), were extracted from the whole-tumor VOI and three habitat subregions using the PyRadiomics v3.0 package. These comprised 93 first-order statistics, 14 shape attributes, 24 gray-level co-occurrence matrix (GLCM) texture features, 16 gray-level run length matrix features, 16 gray-level size zone matrix features, 5 neighboring gray tone difference matrix features, and 14 gray-level difference matrix features, across four sequences. The ICC was used to assess feature repeatability, and only robust features with ICC > 0.75 were retained[19].

Deep learning features: Deep learning models were fed with registered T2WI, ADC, and Ktrans images mapped to three RGB channels. ResNet50, a convolutional neural network pretrained on the ImageNet dataset (available via PyTorch v1.13), was imported. The last fully connected layer was removed, and 2048-dimensional feature vectors were extracted through the global average pooling layer. These were subsequently reduced to 1024 dimensions using principal component analysis to obtain the final deep learning features[20].

Clinical-dosimetric features: Seven features were collected ranging from age, FIGO stage, types of pathology, maximum tumor diameter, squamous-cell carcinoma antigen level to RBE-weighted total dose of ²⁵²Cf neutrons at A point, and overall dose of brachytherapy.

Model building and validation

Feature selection: In the training set, features with variance < 0.01 were first removed. Pearson correlation coefficients were then used to identify and remove highly redundant features (|r| > 0.90). Finally, LASSO regression (sklearn v1.1, solver = "liblinear") was applied for feature selection using 10-fold cross-validation, with the optimal regularization parameter λ determined by the lambda-1se rule. Twenty-three features were selected from 6803 candidate features and incorporated into the final model.

Model training: Feature-layer fusion was performed on the selected 23 features. A support vector machine (RBF kernel, C = 10, γ = 0.01, optimized via grid search combined with 5-fold cross-validation) was trained using binary classification labels: Effective treatment [complete response (CR) + partial response (PR)] and ineffective treatment [stable disease (SD) + progressive disease (PD)].

Comparisons of models: (1) Model-C: Included only the three clinical-dosimetric features (FIGO stage, tumor size, etc.); (2) Model-R: Whole-tumor radiomic features (12 after LASSO selection); (3) Model-RD: Radiomic + deep learning fusion features (17 after LASSO selection); and (4) Model-Fusion: The omni-modal fusion model proposed in this study (23 features).

Validation indicators: Area under the curve (AUC), accuracy, sensitivity, specificity, F1 score; Hosmer-Lemeshow test and calibration curve; clinical net benefit [decision curve analysis (DCA)].

Treatment efficacy evaluation and statistical analysis

Treatment efficacy was evaluated according to RECIST 1.1 criteria based on MRI re-examination at 3-6 months after treatment. Responses were categorized as CR, PR, SD, or PD, with CR + PR defined as effective treatment. Follow-up continued until June 2025. Overall survival (OS) was defined as the time from treatment initiation to death or the last follow-up date. Progression-free survival (PFS) was defined as the time to first progression/relapse or the last follow-up date.

SPSS 26.0 and Python 3.8 (scikit-learn v1.1) were used for statistical analyses. Continuous data were expressed as the mean ± SD, and intergroup comparisons were performed using the independent samples t-test or Mann-Whitney U test. Categorical data, expressed as n (%), were compared using the χ² test or Fisher's exact test. AUC comparisons were performed using DeLong's test. Survival curves were generated using the Kaplan-Meier estimator, and differences between groups were assessed using the log-rank test. A two-sided P < 0.05 was considered statistically significant.

RESULTS

Baseline characteristics of patients

A total of 100 patients were enrolled, with 70 assigned to the training set and 30 to the validation set. The median age was 52 years (range, 31-74 years). Squamous cell carcinoma was the predominant histological type (87 patients, 87.0%), followed by adenocarcinoma (13 patients, 13.0%). According to the FIGO staging system, 17 patients (17.0%) were staged as IB, 41 (41.0%) as IIA-IIB, 30 (30.0%) as IIIA-IIIB, and 12 (12.0%) as IVA. The overall response rate (CR + PR) was 79.0% (79/100) across all patients. No significant differences in baseline characteristics were observed between the training and validation sets (all P > 0.05), indicating that the two groups were comparable (Table 1).

Table 1 Comparison of baseline characteristics of patients in the training set and the validation set.

Characteristic	Training set (n = 70)	Validation set (n = 30)	Statistic	P value
Age (years)	52.3 ± 9.7	51.8 ± 10.2	t = 0.245	0.807
FIGO stage (2018)
IB stage	12 (17.1)	5 (16.7)
IIA-IIB stage	28 (40.0)	13 (43.3)	χ2 = 0.243	0.886
IIIA-IIIB stage	21 (30.0)	9 (30.0)
IVA stage	9 (12.9)	3 (10.0)
Pathological types
Squamous cell carcinoma	61 (87.1)	26 (86.7)	χ2 = 0.003	0.954
Adenocarcinoma	9 (12.9)	4 (13.3)
Maximum tumor diameter (cm)	4.8 ± 1.4	4.6 ± 1.3	t = 0.688	0.493
SCC-Ag (ng/mL; median)	6.2 (2.1-18.4)	5.9 (1.8-20.1)	Z = 0.312	0.755
RBE-weighted total dose of 252Cf neutrons at A point (Gy)	34.2 ± 3.6	33.8 ± 3.9	t = 0.514	0.608
Overall dose of brachytherapy (Gy)	47.8 ± 1.9	48.1 ± 2.0	t = 0.727	0.469
Response to the treatment (efficacy rate)	56 (80.0)	23 (76.7)	χ2 = 0.181	0.671
Complete response	32 (45.7)	14 (46.7)
Partial response	24 (34.3)	9 (30.0)
Stable disease	10 (14.3)	5 (16.7)
Progressive disease	4 (5.7)	2 (6.7)

Measurement data are expressed as the mean ± SD. Enumeration data are expressed as n (%). FIGO: International Federation of Gynecology and Obstetrics; RBE: Relative biological effectiveness; ²⁵²Cf: Californium-252; SCC-Ag: Squamous-cell carcinoma antigen.

Features of habitat maps

Using k-means clustering (k = 3), the tumor VOI was classified into three functional habitat subregions (Table 2): H1 (high vascularity/high cellularity), characterized by high Ktrans (0.312 ± 0.089 min^-1), high Ve, and low ADC, suggesting high proliferative activity and abundant blood supply, accounting for 28.3% ± 11.4% of the tumor volume; H2 (low vascularity/high cellularity), characterized by low Ktrans and low ADC, indicating hypoxia despite high cell density, representing the largest subregion (41.5% ± 13.7%); and H3 (low vascularity/low cellularity), characterized by the lowest Ktrans and high ADC, implying necrosis or low cell density (30.2% ± 14.1%).

Table 2 Features of quantitative parameters in the three habitat subregions.

Habitat subregion	K (per minute)	Ve (mL/mL)	ADC (× 10-3mm2/second)	Proportion of volume (%)	Biological characteristics
H1 (high vascularity/high cellularity)	0.312 ± 0.089	0.487 ± 0.106	0.78 ± 0.15	28.3 ± 11.4	Highly proliferative activity and abundant blood supply
H2 (low vascularity/high cellularity)	0.089 ± 0.031	0.284 ± 0.078	0.69 ± 0.18	41.5 ± 13.7	Hypoxia and high cell density
H3 (low vascularity/low cellularity)	0.062 ± 0.024	0.198 ± 0.065	1.42 ± 0.31	30.2 ± 14.1	Necrosis, hypoxia, or low cell density
P value (H1 vs H2 vs H3)	< 0.001	< 0.001	< 0.001	0.023	-

ADC: Apparent diffusion coefficient.

Correlation analysis revealed that the proportion of H1 was positively correlated with serum squamous cell carcinoma antigen levels (r = 0.38, P < 0.01), suggesting that the H1 subregion reflects active tumor proliferation. Conversely, the proportion of H3 was negatively correlated with the tumor-stroma ratio (r = -0.41, P < 0.01), indicating that the H3 subregion is more closely associated with necrosis or hypoxia. Notably, the proportion of the H3 subregion was significantly higher in the non-responder group (36.4% ± 15.2%) than in the responder group (26.8% ± 12.7%, P = 0.003), suggesting that a higher H3 proportion may be associated with resistance to ²⁵²Cf neutron therapy.

LASSO feature selection

Using LASSO regression for feature selection, 23 features were selected from 6803 candidate features and incorporated into the fusion model (Table 3), including 8 habitat radiomic features (primarily from the H3 subregion), 7 deep learning features (from multi-sequence fusion layers), 5 whole-tumor radiomic features, and 3 clinical-dosimetric features (FIGO stage, RBE-weighted total dose at point A, and maximum tumor diameter). Among these, GLCM entropy (LASSO coefficient 0.312) in the H3 subregion, H3 volume percentage (0.287), and FIGO stage (0.321) were the most important features positively predictive of treatment failure, whereas mean Ktrans (-0.265), RBE-weighted total dose at point A (-0.254), and deep learning feature DL_312 (-0.289) were negative predictors (predictive of effective response) in the H1 subregion.

Table 3 Twenty-three features included in the fusion model after LASSO selection.

Feature style	Feature name	Subregion/sequence	LASSO coefficient	Predictive performance
Habitat radiomics (8)	H3_GLCM_entropy	H3 (low vascularity/low cellularity)	0.312	Ineffective
	H3_ volume proportion	H3/whole-tumor	0.287	Ineffective
	H1_Ktrans_mean value	H1 (high vascularity/high cellularity)	-0.265	Effective
	H2_ADC_skewness	H2 (low vascularity/high cellularity)	0.243	Ineffective
	H1_GLRLM_LRHGE	H1/ADC	-0.198	Effective
	H2/H1 volume ratio	H1 and H2	0.187	Ineffective
	H1_Ve_kurtosis	H1/DCE-MRI	-0.142	Effective
	H3_GLSZM_LZSAE	H3/T2WI	0.131	Ineffective
Deep learning (7)	DL_Feature_312	ResNet50/T2WI	-0.289	Effective
	DL_Feature_578	ResNet50/ADC	0.251	Ineffective
	DL_Feature_049	ResNet50/Ktrans	-0.234	Effective
	The rest 4 features	Multisequence fusion layers	0.09-0.18	-
Whole-tumor radiomics (5)	Whole-tumor_ADC_mean value	DWI/ADC	-0.198	Effective
	Whole-tumor_shape_ sphericity	T2WI	-0.156	Effective
	The rest 3 features	Multisequence	0.08-0.14	-
Clinical-dosimetrics (3)	FIGO stage	Clinical	0.321	Ineffective
	RBE-weighted total dose of 252Cf neutrons at A point	Dosimetrics	-0.254	Effective
	Maximum tumor diameter	Clinical/MRI	0.187	Ineffective

DCE: Dynamic contrast enhanced; MRI: Magnetic resonance imaging; ADC: Apparent diffusion coefficient; T2WI: T2-weighted imaging; FIGO: International Federation of Gynecology and Obstetrics; ²⁵²Cf: Californium-252; RBE: Relative biological effectiveness.

Comparison of prediction model performance

For the independent validation set, the prediction performance of the four models is summarized in Table 4. Model-Fusion achieved an AUC of 0.892 (95%CI: 0.821-0.963), which was significantly superior to Model-C (AUC = 0.712, P = 0.004), Model-R (AUC = 0.783, P = 0.021), and Model-RD (AUC = 0.835, P = 0.038). The accuracy, sensitivity, specificity, and F1 score of Model-Fusion were 0.867, 0.870, 0.857, and 0.863, respectively, demonstrating optimal overall performance (Table 4). Calibration curve analysis (Hosmer-Lemeshow test, χ² = 6.34, P = 0.610) indicated that the predicted probabilities of Model-Fusion were consistent with observed outcomes. DCA showed that across a wide range of threshold probabilities (10%-85%), Model-Fusion consistently provided higher clinical net benefit than the other models (Figure 1).

Figure 1 Model performance comparison in the validation set. A: Receiver operating characteristic (ROC) curves: Model-Fusion (AUC = 0.892) significantly outperformed the other three models (all P < 0.05); B: Calibration curve: Model-Fusion showed good agreement with the ideal diagonal (Hosmer-Lemeshow test, P = 0.610); C: Decision curve analysis: Model-Fusion yielded the highest net benefit across a wide range of threshold probabilities (10%-85%, red shaded area). ROC: Receiver operating characteristic; AUC: Area under curve; TPR: True positive rate; FPR: False positive rate.

Table 4 Comparison of predictive model performance in the validation set.

Model	AUC (95%CI)	Accuracy	Sensitivity	Specificity	F1 score
Model-C (clinical features)	0.712 (0.548-0.876)	0.700	0.652	0.786	0.731
Model-R (whole-tumor radiomics)	0.783 (0.634-0.932)	0.767	0.739	0.810	0.772
Model-RD (radiomics + deep learning)	0.835 (0.701-0.969)	0.833	0.826	0.845	0.823
Model-Fusion (this study)	0.892 (0.821-0.963)	0.867	0.870	0.857	0.863

AUC: Area under curve.

Survival analysis based on the models’ predictive results

Patients with a Model-Fusion prediction probability ≥ 0.50 were classified as the high-benefit prediction group (n = 21), and those with a probability < 0.50 as the low-benefit prediction group (n = 9). Kaplan-Meier analysis (Figure 2) demonstrated that both 3-year PFS (85.7% vs 44.4%, log-rank χ² = 7.038, P = 0.008) and 3-year OS (92.5% vs 71.4%, log-rank χ² = 8.854, P = 0.003) were significantly superior in the high-benefit prediction group compared with the low-benefit group. The median PFS was not reached in the high-benefit group, whereas it was 18.2 months (95%CI: 12.4-24.0 months) in the low-benefit group.

Figure 2 Kaplan-Meier survival curves for different fusion model prediction groups. High benefit group (n = 21, red line) vs low benefit group (n = 9, gray dashed line). A: Progression-free survival (PFS): 3-year PFS 85.7% vs 44.4% (P = 0.008); B: Overall survival (OS): 3-year OS 92.5% vs 71.4% (P = 0.003).

DISCUSSION

In this study, we developed a multimodal predictive model by integrating multiparametric MRI-based habitat imaging with deep learning and radiomics to predict the efficacy of ²⁵²Cf neutron therapy for cervical cancer. In the independent validation set, the model achieved an AUC of 0.892, significantly outperforming single-feature comparison models. Moreover, the model predictions effectively stratified patients according to long-term survival, suggesting considerable potential for guiding clinical practice[21]. Although statistically significant differences in OS and PFS were observed, these findings should be interpreted with caution given the limited validation cohort and the small number of survival events.

The core innovation of this study lies in the introduction of habitat analysis to tumor imaging. Unlike conventional radiomics, which treats tumors as homogeneous wholes, habitat analysis identifies intratumoral multifunctional subregions through unsupervised clustering techniques[5,12]. In our study, H1 (high vascularity/high cellularity) was characterized by high Ktrans and low ADC, representing the tumor core with high proliferative activity and abundant blood supply; H2 (low vascularity/high cellularity) was characterized by low Ktrans and low ADC, corresponding to dense tumor cell clusters within a hypoxic microenvironment; and H3 (low vascularity/low cellularity) exhibited the highest ADC, implying overt necrosis or liquefaction. The H3 subregion may represent a potential target zone where ²⁵²Cf neutron therapy, leveraging its high-LET properties, could overcome the resistance of tumor cells to conventional photon radiotherapy[21]. Furthermore, we found that the proportion of the H3 subregion was significantly higher in patients with ineffective treatment responses. One possible explanation is that the high-LET characteristics of ²⁵²Cf neutron therapy confer a relative advantage in targeting hypoxic tumor cells, partly through the induction of complex DNA damage, including double-strand breaks[22,23]. However, the biological basis of the H3 subregion remains incompletely understood. A larger H3 proportion may reflect differences in the tumor microenvironment or intratumoral heterogeneity that could contribute to treatment resistance[24]. Further studies integrating imaging findings with pathological and molecular analyses are needed to clarify the biological characteristics represented by the H3 subregion.

Recent studies have increasingly recognized habitat imaging as a promising approach for depicting intratumoral heterogeneity. For instance, Cui et al[25] integrated habitat imaging with radiomics and deep learning in a dual-center study of cervical cancer and demonstrated that multimodal integration consistently outperformed individual radiomics or deep learning models in predicting parametrial invasion. Similarly, recent habitat-radiomics research on lymph node metastasis prediction has shown that incorporating habitat-specific information improves predictive accuracy and clinical utility compared with conventional whole-tumor radiomics[26]. These findings support our observation that habitat-derived features, particularly those extracted from the H3 subregion, contribute substantially to treatment-response prediction by capturing biologically meaningful spatial heterogeneity that may not be reflected by whole-tumor analysis alone.

Another important strategy is the fusion of deep learning and radiomic features. Deep learning features can capture complex texture patterns that are difficult for human eyes to discern, whereas radiomic features possess clear physical and mathematical meanings, making them highly interpretable[9-11]. Complementary integration of these two approaches has been shown to yield superior predictive performance compared with either approach alone. Our findings are generally consistent with recent MRI-based artificial intelligence studies in cervical cancer. Jeong et al[17] compared handcrafted radiomics and deep learning models for predicting response to concurrent chemoradiotherapy and demonstrated that both feature types provided complementary predictive information, while integrating clinical variables further improved model performance. Similarly, Cai et al[18] developed a multimodal deep-radiomics model for predicting response to neoadjuvant chemoradiotherapy and achieved an AUC of approximately 0.86 in the validation cohort, highlighting the advantage of combining deep learning representations with handcrafted radiomic features. Compared with these studies, our model further incorporated habitat-derived subregional information, enabling characterization of intratumoral spatial heterogeneity rather than treating the tumor as a homogeneous structure. This additional biological information may partly explain the improved predictive performance observed in our validation cohort (AUC = 0.892), although direct comparisons should be interpreted cautiously given differences in treatment modalities, patient populations, and study endpoints.

Clinically, this model can identify high-risk patients pretreatment who are likely to have poor treatment outcomes (predicted efficacy rate < 50%). For such patients, clinicians could consider intensified strategies, such as moderately increasing the ²⁵²Cf radiation dose, combining with radiosensitizers (e.g., tirapazamine), early assessing the feasibility of surgical downstaging, or introducing ongoing clinical trials of immune checkpoint inhibitor combination therapy, thereby preventing delays caused by ineffective treatment and optimizing the allocation of medical resources. In addition, the continuous prediction probability output can provide quantitative references for shared decision-making between physicians and patients[27].

Several limitations of this study should be acknowledged: (1) This was a single-center retrospective study with a relatively small sample size (100 patients), and external validation in multicenter cohorts is warranted; (2) A fixed k = 3 was used for habitat clustering, and the optimal habitat configuration should be systematically examined in larger cohorts; (3) Although various strategies were adopted during model development, these cannot completely eliminate the possibility of overfitting, particularly given the limited sample size. Future research should incorporate larger-scale multicenter cohorts and external validation datasets to further evaluate the model's robustness, reproducibility, and clinical applicability; (4) Biological interpretation of the habitat subregions requires histopathological spatial verification (e.g., via multiple biopsy immunohistochemistry); and (5) ResNet50 was pretrained on the ImageNet dataset, which may introduce bias during domain migration. Future studies could explore the development of a more specialized model through self-supervised pre-training on large pelvic MRI datasets. Moreover, prospective multicenter validation and integration of multi-omics data (including metabolomics, circulating tumor DNA, etc.) could further strengthen the predictive system.

CONCLUSION

In this study, we successfully constructed and validated a deep learning fusion model based on multimodal MRI habitat mapping for the non-invasive and accurate prediction of ²⁵²Cf neutron therapy efficacy in cervical cancer (AUC = 0.892 in the validation set). The model demonstrated significantly superior performance compared with conventional or single-modality imaging models. This approach provides a promising tool for pretreatment prediction of treatment outcomes to guide personalized chemotherapy and precision radiotherapy for cervical cancer, with considerable potential for clinical translation and broader application.