Radiomics in colorectal cancer patients

Abstract

The main therapeutic options for colorectal cancer are surgical resection and adjuvant chemotherapy in non-metastatic disease. However, the evaluation of the overall adjuvant chemotherapy benefit in patients with a high risk of recurrence is challenging. Radiological images can represent a source of data that can be analyzed by using automated computer-based techniques, working on numerical information coded within Digital Imaging and Communications in Medicine files: This image numerical analysis has been named “radiomics”. Radiomics allows the extraction of quantitative features from radiological images, mainly invisible to the naked eye, that can be further analyzed by artificial intelligence algorithms. Radiomics is expanding in oncology to either understand tumor biology or for the development of imaging biomarkers for diagnosis, staging, and prognosis, prediction of treatment response and diseases monitoring and surveillance. Several efforts have been made to develop radiomics signatures for colorectal cancer patient using computed tomography (CT) images with different aims: The preoperative prediction of lymph node metastasis, detecting BRAF and RAS gene mutations. Moreover, the use of delta-radiomics allows the analysis of variations of the radiomics parameters extracted from CT scans performed at different timepoints. Most published studies concerning radiomics and magnetic resonance imaging (MRI) mainly focused on the response of advanced tumors that underwent neoadjuvant therapy. Nodes status is the main determinant of adjuvant chemotherapy. Therefore, several radiomics model based on MRI, especially on T2-weighted images and ADC maps, for the preoperative prediction of nodes metastasis in rectal cancer has been developed. Current studies mostly focused on the applications of radiomics in positron emission tomography/CT for the prediction of survival after curative surgical resection and assessment of response following neoadjuvant chemoradiotherapy. Since colorectal liver metastases develop in about 25% of patients with colorectal carcinoma, the main diagnostic tasks of radiomics should be the detection of synchronous and metachronous lesions. Radiomics could be an additional tool in clinical setting, especially in identifying patients with high-risk disease. Nevertheless, radiomics has numerous shortcomings that make daily use extremely difficult. Further studies are needed to assess performance of radiomics in stratifying patients with high-risk disease.

Key Words: Colorectal cancer, Radiomics, Artificial intelligence, Liver metastases, Magnetic resonance imaging, Computed tomography, Positron emission tomography/computed tomography

Core Tip: Stratifying colorectal cancer patients with high-risk disease and the evaluation of the overall chemotherapy benefit are a clinical challenge. Radiomics through radiological images analysis using automated computer-based techniques allows the extraction of quantitative features from radiological images, mainly invisible to the naked eye, that can be further analyzed by artificial intelligence algorithms. Several efforts have been made to develop radiomics signatures for colorectal cancer patient using computed tomography (CT), magnetic resonance imaging, and positron emission tomography/CT, in particular to understand tumor biology, to develop imaging biomarkers for diagnosis, staging, and prognosis, to predict treatment response and to monitor disease.

INTRODUCTION

Colorectal cancer is the fifth-most-common frequent in terms of incidence and mortality, with 1480000 new cases in 2020 worldwide[1]. The TNM staging process itself is widely based on radiological definition of boundaries of primary lesion, nodal and distant metastases. The main therapeutic options for colorectal cancer are surgical resection and adjuvant chemotherapy in non-metastatic patients; however, the evaluation of the overall adjuvant chemotherapy benefit in patients with a high risk of recurrence is a clinical challenge[2]. The decision is based on the TNM staging system[3], which represents the most important parameter: Colorectal cancer patients at stage III are globally recognized as patients who can benefit from chemotherapy, while for those at stage II with other clinical risk factors, the advantages of chemotherapy are still debated[2,4]. In presence of clinical risk factors, the final strategy is often decided by the oncologist or multidisciplinary teams.

Nowadays radiological images can represent a source of data that can be analyzed by using advanced computer-based techniques, working on numerical information coded within the Digital Imaging and Communications in Medicine files[5]: This image numerical analysis has been named “radiomics”[6]. Radiomics might be used as a non-invasive imaging biomarker and be able to provide a quantitative evaluation of medical images, with the chance to shift imaging from a qualitative to a quantitative approach[7,8]. To date, the radiomics approach has been extensively investigated in cancer patients with a specific focus on tumor diagnosis, staging, prognosis prediction, and long-term monitoring[7,9,10]. In this context, radiomics could play a pivotal role in colorectal cancer workup with the expectancy to help clinicians in identifying patients with high-risk disease.

RADIOMICS APPLIED TO CANCER-SUMMARY (PATHOLOGICAL CORRELATION, DIAGNOSIS, PROGNOSIS)

Since its advent in 2012[5], radiomics has been extensively applied in oncology studies and it has been demonstrated as a promising tool that can offer a risk-free and efficient method for diagnosis, classification, and prognosis prediction in oncology[11]. Radiomics, indeed, allows the extraction of quantitative features from radiological images, mainly invisible to the naked eye, that can be further analyzed by machine learning and artificial intelligence (AI) algorithms to produce signatures representing tumor phenotype.

The radiomics pipeline consists of different steps (Figure 1). First, multi-dimensional and multi-institutional data should be collected including high-quality medical images, clinical, and, eventually, molecular data. Once acquired, images should be pre-processed to improve image quality, harmonize raw data, and ensure generalizability across imaging protocols and patient populations, especially if multicentre datasets are acquired. This pre-processing step usually involves image co-registration, image denoising, signal intensity standardization, and/or normalization. Once the image datasets have been pre-processed, tumors should be segmented to extract regions of interest (ROIs) on which subsequent steps will be focused on. This task can be performed either manually, semi- or fully automatically. A big effort in the research field is addressed to the development of AI-based systems to automatically segment lesions and overcomes the most common limitations of manual segmentation. Afterward, from the segmented ROIs a large number of parameters (features) are extracted, including: (1) First-order features, from gray-level intensity histograms and lesion shape; (2) second-order features, related to the spatial relationship between pixels, calculated using different matrices, e.g., gray-level co-occurrence, gray-level run-length, gray-level dependence, gray-level size zone, neighboring gray tone difference; and (3) transform-based features, e.g., Wavelet, Gabor, Laws, Laplacian. However, since the number of extracted features could be much larger than the sample size of patients included in the algorithms’ development, it is vital to reduce the number of features through a step called features selection. This step will strongly reduce the risk of overfitting[12], which occurs when the algorithm overadapts its performances based on data in the training set and consequently loses its generalizability. Besides, features selection will ensue to exclude features that are non-reproducible, redundant, and/or non-relevant for the task, and to reduce the computational cost, while improving the performance of the model[13]. Once the most performing features are selected, the radiomics signature is finally developed by using algorithms for classification, such as logistic regression[14,15], k-nearest neighbour[16], naïve Bayes classifier[17], support vector machines (SVM)[18,19], random forest (RF)[20,21], neural network[22,23] and deep learning[24,25]. In this step, is crucial to divide the image dataset into three subgroups: One for training, used to develop the algorithm, the second for testing, to fine-tune the model, and the last for validation, which aims to evaluate the performance on a different dataset. Training and testing could be performed also using a cross-validation approach, i.e., in which different portions of the dataset are iteratively used to train and test the model. Conversely, the validation should be performed using patients that were never seen during the development of the algorithms. The validation step can be internal, when applied in a similar clinical setting and population to the training set, or preferably external, when applied in multiple clinical settings with varying disease prevalence[12,26].

Figure 1 Radiomics pipeline. ROI: Regions of interest.

Thanks to the vast number of images routinely used by radiologists and oncologists in their daily workload, radiomics is substantial in oncology to understand tumor biology or develop imaging biomarkers useful for diagnosis, staging, prediction of treatment response, disease monitoring and surveillance[27].

Understanding tumor biology through radiomics can be feasible because it allows the extraction of quantitative information about spatial and temporal heterogeneity in a non-invasively way and using routinely acquired images. This information can be consequently correlated with tumors’ phenotype that can either reflects distinct traits (e.g., internal necrosis and proliferation at the periphery) or mirrors genomic and molecular traits or be a signature or different outcomes. Moreover, texture features used in radiomics have been also demonstrated useful in reflecting key oncogenomics processes such as tumor angiogenesis[28], hypoxia[29], tumor invasion[30] and tumor proliferation[31].

The second scope of radiomics in the oncology field is to enhance precision medicine through the implementation of diagnostics and prognostic imaging biomarkers in a variety of solid tumors. Biomarkers for detection and diagnosis are those in a more advanced status since there are many studies that demonstrated their usefulness in discriminating between healthy, benign and malignant cancer in different sites[23,32-34]. However, the most promising applications in which radiomics could truly improve clinical practice are related to the prediction of treatment response and disease monitoring. Indeed, knowing, before or during therapy, which patients would respond might help choosing the best management possible. Moreover, after treatment, radiomics biomarkers may suggest more intense post-treatment surveillance due to a high risk of a tumor recurrence for a particular patient. In parallel to radiomics, it is worthwhile underling that a boost for the development of prognostic biomarkers for precision medicine could be provided by integrating radiomics features to additional layers of -omics information, i.e., pathomics (features derived from digital pathological samples), and genomics. The motivation for this multi-omics approach to disease understanding is that the conventional markers discovery which molecularly dissect the disease part by part, if the sum of knowledge of parts will explain the operation of the whole, has mostly failed to understand the causes and cures for complex diseases. On the contrary, recent evidence suggests that patterns discovered from high dimensional, multi-modal data could improve estimation of disease aggressiveness and patient outcomes compared to single modality data.

COMPUTED TOMOGRAPHY

Tumor and nodes

Several efforts have been made to develop radiomics signatures for colorectal cancer patient using computed tomography (CT) images with different aims (Table 1). Li et al[35] developed and validated a clinical-radiomics nomogram for the preoperative prediction of nodes metastasis. They validated their algorithm on an internal dataset of 308 patients (136 with and 172 without lymph node metastases) and showed that the model which included clinical parameters, radiomics on both tumor and peripheral nodes was the one reaching the highest accuracy in predicting nodes metastases [area under the curve (AUC) = 0.7509; 95%CI: 0.6901-0.8071; accuracy: 73.70%; sensitivity: 60.29%; specificity: 84.30%; positive predictive value (PPV): 75.23%; and negative predictive value (NPV): 72.86% in the internal validation set]. If further validated, also on an external validation set, this model could be used as an individualized preoperative non-invasive tool, assisting in clinical treatment decision making and achieving precision treatment.

Table 1 Summary of the most important published papers regarding the usefulness of radiomics in colorectal cancer patients using computed tomography imaging.

Ref.	Imaging	Main aim	Patients (n)	Main findings
Li et al[35], 2020	CT	Prediction of nodes metastases	766	Overall diagnostic values: Sensitivity = 60.3%; specificity = 84.3%; PPV = 75.2%; NPV = 72.9%; AUC = 0.750
Shi et al[16], 2020	CT	Detect RAS and BRAF phenotypes	159	Combined score (semantic features and radiomics) AUC = 0.950; validation cohort AUC = 0.790
Giannini et al[41], 2020	CT	Predict response to treatment	38 (141 lesions)	Per-lesion diagnostic values: Sensitivity = 89%; specificity = 85%; PPV = 78%; NPV = 93%
Dercle et al[47], 2020	CT	Tumor response to anti-EGFR therapy	667	Sensitivity to therapy: AUCs 0.800 and 0.720 for FOLFIRI and FOLFIRI + cetuximab
Dohan et al[48], 2020	CT	Overall survival	491	SPECTRA score > 0.02 has a lower OS; SPECTRA Score at 2 mo has the same prognostic values as RECIST at 6 mo
Giannini et al[41], 2020	CT	Predict response to treatment	57 (242 lesions)	Per-lesion diagnostic values: Sensitivity = 99%; specificity = 94%; PPV = 95%; NPV = 99%; the radiomic approach can predict R- wrongly classified by RECIST as R+
Taghavi et al[103], 2021	CT	Prediction of synchronous liver metastases	91	The radiomics model outperformed the clinical model: AUC = 0.93 vs 0.64
Rao et al[108], 2014	CT	Prediction of synchronous liver metastases	29	The mean entropy of the liver is significantly higher in metastatic patients (P = 0.02); Liver entropy can help the differential between metastatic and non-metastatic patients (AUC = 0.73-0.78)
Li et al[109], 2022	CT	Prediction of synchronous liver metastases	323	A combined clinical-radiomics model has a good AUC (= 0.79) in detecting liver metastases
Ng et al[111], 2013	CT	Prediction of overall survival	55	Entropy, uniformity, kurtosis, skewness, and standard deviation of the pixel distribution histogram can predict survival; each parameter can be considered an independent predictor of the overall survival state
Mühlberg et al[112], 2021	CT	Prediction of overall survival	103	Tumor burden score can discriminate patients with at least 1-year survival (AUC = 0.70); a machine-learning model better predict survival (AUC = 0.73)
Ravanelli et al[116], 2019	CT	Prediction of response and prognosis after chemotherapy	43	Uniformity is lower in responders (P < 0.001); uniformity is independently correlated with radiological response (OR = 20.00), overall survival (RR = 6.94) and progression-free survival (RR = 5.05)

PPV: Positive predictive value; NPV: Negative predictive value; AUC: Area under the curve; OS: Overall survival; SPECTRA: Survival PrEdiction in patients treated by FOLFIRI and bevacizumab for mCRC using contrast-enhanced computed tomography TextuRe Analysis; CT: Computed tomography.

From another point of view, radiomics has also been proven effective in detecting BRAF and RAS (KRAS and NRAS) gene mutations, that are genomics signatures usually associated with shorter disease-free and overall survival. These mutations are determined through genetic molecular profiling by sampling the tumor, however biopsy carries several drawbacks, including the risk of adverse events, such as bleeding, physical and psychological discomfort[36].

For these reasons, radiomics could potentially be used to non-invasively predict RAS and BRAF mutation status in patients with colorectal cancer and to further guide treatments with surgery or chemotherapy[16,37]. Shi et al[16] validated a combined score that tracks RAS (KRAS and NRAS) and BRAF mutant phenotypes in colorectal cancer from multicentre CT image data AUC of 0.79 on the validation cohort.

Recently, efforts have been made to translate radiomics signatures from a patient level to a lesion level, since it has been demonstrated that heterogeneous response, caused by the onset of new resistant tumor clones in some lesions, is a predictor of poor overall survival[38-40]. Differentiating which colorectal liver metastases (CRLM) responds and which lingers and eventually will progress in the same patient could pave the way to truly personalized treatment. Giannini et al[41] preliminary demonstrated the feasibility of using radiomics features from baseline CT to predict response of treatment after 3 mo. They validated the signature on an independent cohort of patients obtaining encouraging results especially in identifying patients with outlier lesions, i.e., that do not respond in general condition were most lesions respond. In these cases, a target biopsy on non-responder lesions could have revealed a different genetic makeup or, in absence of extrahepatic lesions, suggested the local ablation of outlier metastases. More recently, another breakthrough has been made using the use of delta-radiomics, whose aim is to assess the treatment-induced change of radiomics features over time that could provide information about prognosis[42,43]. These variations can be measured in different ways, for example as the differences between features computed on the same tumour before and after treatment[44,45] or the net-change (i.e., difference of radiomics features after treatment over the value before treatment)[46]. Other than providing additional information about tumour behaviour, delta-radiomics represents a very interesting approach since it could theoretically allow to adapt and modulate the ongoing treatment approach thanks to the predictive power of this technique[42].

Delta-radiomics has been already proven effective in predicting overall survival in patients with metastatic colorectal cancer[47,48]. Dercle et al[47] developed and validated on a multicentre dataset a delta-radiomics associated with tumor sensitivity to anti-EGFR therapy in colorectal cancer patients (AUC = 0.80). Similarly, Dohan et al[48] validated a delta-radiomics signature able to predict overall survival and identify good responders better than RECIST1.1 criteria in patients with metastatic colorectal cancer treated by FOLFIRI and bevacizumab as a first-line treatment. From a per-lesion point of view, Giannini et al[45] validated a delta-radiomics signature able to predict long-term response (i.e., more than 8 mo) of individual CRLM with an accuracy of 86% in the validation dataset. Of note, the delta-radiomics signature was able to reliably predict non-responder liver metastases wrongly classified as responder by lesion RECIST at the first time point. This per-lesion approach could strongly impact treatment, since according to the delta-radiomics signature it would be possible to pinpoint lesions with distinct biological and molecular features, thus enabling studies toward lesion-specific personalized treatment in liver-only metastatic colorectal cancer patients.

MAGNETIC RESONANCE IMAGING

Tumor evaluation

The application of radiomics and texture analysis by using magnetic resonance imaging (MRI) images has increased interest in recent years (Table 2), with a specific focus on liver pathology, renal carcinoma, prostate cancer, and, in slight minority of published studies, rectal cancer[49]. As mentioned above, the staging of rectal cancer is mainly based on MRI; however, radiological images analyzed by dedicated software can add important data useful for the best management of patients. In this setting, it is of utmost importance to underline the potentiality of radiomics as a non-invasive biomarker for predicting histopathological data, as demonstrated for different abdominal pathological conditions, related to the liver, pancreas, and colorectal[8]. Even if for many abdominal organs it can be difficult to obtain a useful histological sample, rectal cancer pathological data are easy to collect, considering that colonoscopy or sigmoidoscopy, depending on the location of the lesion, is the reference standard technique[50]. On these bases, most published studies when this search was performed mainly focused on the response of advanced tumors that underwent neoadjuvant therapy.

Table 2 Summary of the most important published papers regarding the usefulness of radiomics in colorectal cancer patients using magnetic resonance imaging.

Ref.	Imaging	Main aim	Patients (n)	Main findings
Horvat et al[52], 2022	MRI	Response to chemotherapy	114	Combined radiological-radiomics model increased agreement (κ = 0.82 vs κ = 0.25)
Dinapoli et al[53], 2018	MRI	Pathological complete response	221	Significant covariates, skewness, and entropy can predict pathological complete response, with AUCs = 0.730 and 0.750 for internal and external cohorts
Shahzadi et al[50], 2022	MRI	Response to chemotherapy	190	Radiomics combined with the T stage better predict response
Liu et al[23], 2021	MRI	Prediction of nodes metastases	186	Clinical-radiomics model improves performance: AUC = 0.827
Chen et al[72], 2022	MRI	Tumor differentiation and nodes metastases	37 (487 nodes)	Radiomics features of the primary tumor can predict tumor differentiation: AUC = 0.798
Liu et al[73], 2017	MRI	Tumor differentiation	68	Skewness and entropy are lower in pT1-2 in comparison with pT3-4 (P < 0.05)
Yang et al[74], 2019	MRI	Prediction of T and N stage	88	Skewness, kurtosis, and energy are higher in metastatic nodes in comparison with non-metastatic ones (P < 0.001)
Ma et al[75], 2019	MRI	Prediction of nodes metastases and N staging	152	SVM has higher diagnostic values for T and N stages (AUC = 0.862) in comparison with MLP and RF
Zhu et al[76], 2019	MRI	Prediction of nodes metastases	215	Radiomic model AUC = 0.818
Zhou et al[77], 2020	MRI	Prediction of nodes metastases	391	The combined model predicts nodes metastases: NPV = 93.7%, AUC = 0.818
Shu et al[34], 2019	MRI	Prediction of synchronous liver metastases	194	The Radiomics model combined clinical risk factors and LASSO features and showed a good predictive performance: AUC = 0.921
Liu et al[107], 2020	MRI	Prediction of synchronous liver metastases	127	A radiomic nomogram presents an accuracy of 81.6% in predicting liver metastases (AUC = 0.918)
Granata et al[115], 2022	MRI	Prediction of overall survival	90	Second-order features can predict infiltrative tumor growth, tumor budding, and mucinous type; a second-order feature can predict the risk of recurrence with an accuracy of 90%
Jalil et al[119], 2017	MRI	Prediction of prognosis after chemotherapy	56	MPP can predict overall survival (HR = 6.9) and disease-free survival (HR = 3.36); texture analysis can predict relapse-free survival on pre- and post-treatment analyses

AUC: Area under the curve; SVM: Support vector machine; MLP: Multilayer perceptron; RF: Random forest; NPV: Negative predictive value; HR: Hazard ratio; MPP: Mean positive pixel; MRI: Magnetic resonance imaging.

Recently, Chen et al[51], in a single-center prospective study, enrolled 137 patients who underwent neoadjuvant chemotherapy. The Authors demonstrated that the traditional clinical model reported an AUC of 67.6% and 70.1% in the training and validation cohort, respectively, quite similar to the selective clinical model (77.5% and 59.6%, respectively). On the other hand, when combining radiomics with clinical data the AUC raised to 94.9% and 84.4% in the training and validation cohort, respectively.

Similarly, Horvat et al[52], by enrolling 114 patients who underwent neoadjuvant chemotherapy, demonstrated that radiomics can help the radiologist determine the pathological complete response. The Authors found that combined clinical and radiomics models increased the agreement compared with radiologist interpretation and can help the less experienced radiologist in increasing diagnostic values, in particular specificity, PPV, and NPV.

Analogously, Dinapoli et al[53] analyzed the radiomics data of 221 patients from three different centers and demonstrated that this tool can help the prediction of pathological complete response before starting neoadjuvant chemotherapy. Moreover, the Authors performed an external validation, to test the obtained results, with good diagnostic values.

Even if few studies are published in the literature, to test the robustness of the radiomics approach in rectal cancer patients, Shahzadi et al[54], demonstrated that only one study can be used for external validation, underlying the overall lack of reproducibility and the need of further standardization before considered it a useful clinical tool. In this setting, future directions should be focused on multicentre studies with standardized MR protocols to validate and test the feasibility of the radiomics approach and its potential usefulness in current everyday clinical practice.

Node’s evaluation

Nodes metastases is the main metastatic site of colorectal cancer and an important cause of post-operative recurrence and death[55]. Nodes status is a key factor in the TNM staging of colorectal cancer and the main determinant of adjuvant chemotherapy[56,57].

Preoperative knowledge of NS can provide valuable information for determining the need for adjuvant therapy and the adequacy of surgical resection, thus aiding in pre-treatment decision making[58].

In clinical practice, CT is the most used preoperative imaging method to detect metastatic lesions and perform tumor staging in patients with colorectal cancer. However, the limitation of CT examination is that it cannot discriminate between benign and malignant nodes[59].

MRI has the highest contrast resolution for soft tissues, allowing the best depiction of neoplastic lesions, their anatomical relationships, the depth of the rectal wall involvement, extramural venous invasion, circumferential resection margins, and the assessment of the N stage. For these reasons, MRI examination is considered the reference standard for locoregional staging and restaging in RC according to the main international guidelines[60-62].

Advances in pattern recognition tools and the increase in data set sizes have facilitated the development of radiomics, which may potentially improve predictive accuracy in oncology[63].

Therefore, in the last decade, several papers have been published, with different imaging techniques, reporting the potential role of radiomics in diagnosis, characterization, and evaluation of the tumor response to treatments[64-67] and nodal assessment[68-70].

MRI can provide multiparameter images different from those obtained by CT, so it is of interest whether there exists an association between NS and multiregional radiomics features of multiparametric MR images in rectal cancer patients[71].

Liu et al[71] aimed to develop and validate a multiregional radiomics prediction model based on MRI and combine it with clinical-semantic data for the individualized preoperative prediction of lymph node metastasis in rectal cancer patients.

Similarly, the study of Chen et al[72] provides two non-invasive and quantitative methods, which respectively predict the tumor differentiation and regional nodes metastases for rectal cancer preoperatively. MRI images of both the primary tumor alongside the lymph nodes and specimens were performed with a node-to-node match and labeling. A prediction model was then successfully developed, which provided AUC values of 84.6% and 73.3% in the training and test cohort, respectively.

Liu et al[73] performed a histogram analysis on the ADC map of the whole tumor and reported that entropy was an independent predictor of nodal involvement. Recently, Yang et al[74] performed the same analysis on T2-weighted imaging of the whole tumor: They found that a lower skewness was an independent risk factor for lymph node metastases.

In a recent retrospective single-center study, radiomics features were extracted from preoperative high-resolution T2-weighted imaging of different histological RC and analyzed using different algorithms. The RF analysis showed a good diagnostic performance for the N-stage with an AUC of 74.6%. The prediction model was able to differentiate N0 from N1-N2 patients with a sensitivity of 79.0% and a specificity of 72.0%[75].

Zhu et al[76] compared the performance of two models based, respectively, on the radiomics signature of the primary tumor and of the lymph nodes, before and after chemoradiotherapy (CRT), for the prediction of nodal involvement in advanced rectal cancer. The authors concluded that the features from the lymph node model perform better than the tumor features for the prediction of nodal involvement[76].

Similarly, Zhou et al[77] evaluated a multi-parametric MRI radiomics model for nodal assessment following CRT by combining the radiomic signature with an experienced radiologist’s visual evaluation: stratified analyses indicated that the combined model could predict lymph node metastasis with a NPV of 100 and 87.8% after treatment[77].

Even if current literature is focusing on the importance and applicability of radiomics in rectal cancer, no significant studies have been published in the field of colorectal cancer, mainly because MR is not still validated as imaging technique for local staging and restaging of this kind of tumor.

PET/CT

18F-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) is frequently employed in the staging and post-neoadjuvant assessment of patients with rectal cancer. PET/CT provides functional information of the primary tumor, nodes and distant metastases. In patients with colorectal cancer, the combination of texture parameters with the functional information obtained with PET/CT scan can further enhance the predictive power of PET/CT imaging[78]. Current studies mostly focused on the applications of radiomics in PET/CT for the prediction of survival after curative surgical resection and assessment of response following neoadjuvant chemoradiotherapy (Table 3).

Table 3 Summary of the most important published papers regarding the usefulness of radiomics in colorectal cancer patients using positron emission tomography/computed tomography imaging.

Ref.	Imaging	Main aim	Patients (n)	Main findings
Lovinfosse et al[80], 2018	PET/CT	Progression-free and overall survival	86	SUVmean, dissimilarity, and contrast from the neighborhood intensity-difference matrix are independently associated with overall survival
Hotta et al[81], 2021	PET/CT	Progression-free and overall survival	94	MTV, TLG, and GLCM entropy are associated with overall survival; SUVmax, MTV, TLG, and GLCM entropy are associated with progression-free survival
Bundschuh et al[83], 2014	PET/CT	Response after neoadjuvant chemotherapy	27	COV can assess histopathologic response during (sensitivity 68%, specificity 88%) and after (sensitivity 79%, specificity 88%) therapy
Bang et al[84], 2016	PET/CT	Response after neoadjuvant chemotherapy	74	MV is associated with 3-yr disease-free survival; Kurtosis and kurtosis gradient are associated with 3-yr disease-free survival
Giannini et al[85], 2019	PET/CT	Response after neoadjuvant chemotherapy	52	Second-order texture features (five from PET and one from MRI) can help distinguish responder and non-responder patients: Sensitivity = 86%; specificity = 83%; AUC = 0.860
Yuan et al[89], 2021	PET/CT	Response after neoadjuvant chemotherapy	66	A radiomics model can predict TRG 0 vs TRG 1-3: Sensitivity = 77.8%, specificity = 89.7%, AUC = 0.858
Schurink et al[86], 2021	PET/CT	Response after neoadjuvant chemotherapy	61	Combined baseline and global tumor features better predict response compared to baseline and local texture (AUC = 0.83 vs 0.79)
Shen et al[87], 2020	PET/CT	Predict pathological complete response	169	RF can predict complete response: Sensitivity = 81.8%; specificity = 97.3%; PPV = 81.8%; NPV = 97.3%; accuracy = 95.3%
He et al[90], 2021	PET/CT	Prediction of nodes metastases	199	Logist regression and XGBoost can accurately predict nodes metastases with AUC = 0.866 and 0.903, respectively
Ma et al[91], 2022	PET/CT	Prediction of perineural invasion and outcome	131	12 radiomics signatures are associated with peri-neural invasion; a radiomic score can differentiate between perineural positive and negative lesions: AUC = 0.900
Li et al[92], 2021	PET/CT	Prediction of microsatellite instability	173	2 radiomics features can predict microsatellite instability: Sensitivity = 83.3%; specificity = 76.3%; accuracy = 76.8%
Lovinfosse et al[93], 2016	PET/CT	Prediction of RAS status	151	SUVmax, SUV mean, skewness, SUV standard deviation, and SUV coefficient of variation are associated with RAF mutation (all P < 0.001)
Chen et al[94], 2019	PET/CT	Prediction of genetic mutations	74	MTV and SUV max are increased in mutated KRAS tumors (all P < 0.001); short-run low gray-level emphasis is associated with p53 mutations (P = 0.001); gray-level zone emphasis is associated with APC mutations (P = 0.006)

PET/CT: Positron emission tomography/computed tomography; PPV: Positive predictive value; NPV: Negative predictive value; AUC: Area under the curve; RF: Random forest; MTV: Metabolic tumor volume; TLG: Total lesion glycolysis; GLCM: Gray-level co-occurrence matrix; COV: Coefficient of variation; MV: Metabolic volume; OR: Odd ratio; MPP: Mean positive pixel.

For the prediction of prognosis in patients with colorectal cancer, Kang et al[79] showed that radiomics score from baseline PET scans was significantly associated with progression-free survival. Similarly, Lovinfosse et al[80] and Hotta et al[81] correlated the texture features in PET/CT with both progression-free survival and overall survival. Furthermore, a recent study provided a combined clinical-radiomics model with high predictive performance (C-index of 0.780) for recurrence-free survival in 196 patients with PET/CT[82].

Several studies explored the potential of radiomics for the prediction of response and survival after neoadjuvant chemoradiotherapy[83-89]. In an initial study of 27 patients with rectal cancer treated with neoadjuvant chemoradiotherapy, Bundschuh et al[83] calculated texture parameters (skewness and kurtosis) on PET/CT, which provided a good performance for late response prediction but no significant predictive capability for the assessment of early response. In a retrospective study performed by Bang et al[84], texture parameters extracted from PET images correlated with both tumor regression grading and disease-free survival in patients with rectal cancer undergoing neoadjuvant chemoradiotherapy. Giannini et al[85] combined radiomics features from PET and MRI to predict pathological complete response following neoadjuvant chemoradiotherapy with high accuracy (AUC of 0.86) in patients with rectal cancer. Similarly, Schurink et al[86] combined pretreatment tumor features on PET/CT and MRI to predict response to chemoradiotherapy in rectal cancer, with an AUC of 0.81. Shen et al[87] developed a RF model to predict pathological complete response after neoadjuvant chemoradiotherapy in 169 patients with rectal cancer, which demonstrated a sensitivity of 97.3% and a specificity of 81.8% for the identifications of cancers with complete response. Nevertheless, a study by Karahan Şen et al[88] found no superiority of texture features compared to metabolic tumor volume in predicting response to neoadjuvant chemoradiotherapy in patients with locally advanced rectal cancer.

For the prediction of lymph node metastasis, a retrospective study published by He and colleagues[90] analyzed the radiomics score and five machine-learning models to predict metastatic lymph nodes based on the radiomics features of 199 colorectal cancers, with an AUC of 0.747-0.581 in the test set. Additionally, the performance of PET/CT radiomics features for predicting perineural invasion has been explored in a recent retrospective study[91].

Finally, few other studies explored the correlation between radiomics signature extracted from PET/CT and rectal cancer genotypes, such as microsatellite instability status[92], RAS mutational status[93,94], TP53 and adenomatous polyposis coli mutations[94].

Despite the promising value of radiomics in PET/CT scans of rectal cancer, it should be noted that all current results are based on retrospective single-center studies with heterogeneity on the type of extracted features and analysis. Moreover, several current studies demonstrated insufficient quality according to the radiomics quality score assessment[95].

CRLM

CRLM develop in about 25% of patients with colorectal carcinoma, being more commonly synchronous (14%-17%) rather than metachronous (8%-15%)[96-99]. CT is most adopted to detect CRLM at preoperative staging due to its higher availability compared to MRI, while MRI is usually used in selected doubtful cases particularly in the challenging scenario of the “too small to characterize” hypoattenuating lesion. In addition, chemotherapy regimens may cause focal or diffuse hepatic changes at imaging that can profoundly alter visualization of hepatic metastases on CT, reducing its diagnostic accuracy and MRI proves to be helpful as problem-solving tool in some cases[100]. Therefore, the overall sensitivity, specificity and accuracy for the diagnosis of CRLM are lower for CT compared to MRI[100-102]. For this reason, at some center, abbreviated gadoxetate disodium MRI protocols are adopted rather than trusting CT only[103,104].

The adoption of radiomics has been proven successful in diagnostic, prognostic, and therapeutic stages[78].

Diagnosis and risk assessment

In term of diagnosis of synchronous classical logistic regression models (CLRM), it is relevant to highlight that even with MRI the sensitivity may be lower than 80%, particularly in patients with mucinous adenocarcinoma as primary tumor, prior local treatment in the liver or metastases smaller than 1 cm[102]. Therefore, the first main diagnostic task of radiomics should be the identification of CRLM before they can be seen by radiologist’s naked eye (i.e., detection of synchronous metastases). In a pilot study, Devoto et al[105] proved that radiomics can potentially predict the development of liver metastases on baseline liver CT, by demonstrating a higher heterogeneity of liver texture analysis in patients who developed liver metastases compared to patients who did not develop them. Other authors[34,106] investigated whether radiomics applied to T2-weighted images of the primary tumor on MRI could help in the preoperative prediction of CRLM: Shu et al[34] used a region of interest while Liu et al[107] used a volume of interest and both demonstrated that a radiomics nomogram constructed by combining radiomics and clinical data achieve AUC higher than 90% in the preoperative prediction of CRLM.

The second main diagnostic task of radiomics should be preoperative identification of patients at risk of developing CLRM (i.e., detection of metachronous metastases) based on micro-environmental changes in the apparently normal liver. Taghavi et al[106] and Rao et al[108] designed a prediction model based on liver CT radiomics for the detection of metachronous CRL, with the first study including more patients imaged at three centers and combining radiomics and clinical data achieving AUC of up to 86% in the validation cohort. Other studies tried to achieve the same goal by assessing the primary tumor on CT[109] or MRI[110]. Specifically, Li et al[109] obtained an AUC of 0.72 for the prediction of metachronous CLRM by combining clinical data and volumetric radiomics of the primary tumor on CT. In regard of MRI, a systematic review including 1497 patients estimated a pooled sensitivity and specificity of radiomics applied to rectal MRI of 0.76 and 0.85 respectively in predicting metachronous CLRM, and AUC of the included studies ranging from 0.83 to 0.87[110].

Prognosis

In terms of prognostic information, radiomics of CRLM has emerged as a promising tool to preoperatively predict patient survival at diagnosis and after therapy. Ng et al[111] suggested that tumors demonstrating less texture tumor heterogeneity using radiomic CT analysis may predict poorer survival at diagnosis[111]. Mühlberg et al[112] showed that CT-based geometric distribution and radiomics analysis of whole liver tumor burden from preoperative CT may help for prediction of 1-year survival. Radiomics of CRLM on CT seems also promising for differentiating desmoplastic from replacement histopathological growth patterns[113], and this differentiation may provide an earlier estimate of disease aggressiveness and prognosis as the desmoplastic histopathological growth pattern usually has longer overall survival[114]. As demonstrated by Granata et al[115] contrast MR-based radiomics and machine learning analysis may help in the preoperative prediction of the front of tumor growth (expansive or infiltrative), the tumor budding (absent, low grade or high grade) and tumor recurrence after surgery, all of which may affect patient outcome. Studies looking at survival after chemotherapy in patients with CRLMs obtained similar results regarding the role of texture homogeneity/heterogeneity in the prediction of prognosis[116,117]. As an example, Ravanelli et al[116] demonstrated that lower uniformity of CRLM on CT texture analysis was independently correlated with overall survival and progression free survival in patients treated with bevacizumab, but not in those treated with standard chemotherapy. Other radiomic features such as entropy, kurtosis, and skewness have been investigated so far on CT and MRI, all providing an additional piece of the puzzle and supporting the concept that the addition of texture analysis in the pre-treatment assessment may provide information on prognosis in patients with primary colorectal cancer and CLRM[118,119].

Treatment response

Finally, the correct assessment of response in the treatment of CLRM and the prompt prediction of early response is of utmost important in defining the success or failure of treatment interventions and in the selection of those patients requiring a change of the therapeutic regimen. Chemotherapy for CLRM in the modern era of oxaliplatin- and irinotecan-containing regimens (e.g., FOLFOX, FOLFIRI, CAPOX/FOLFOXIRI, XELOX) has been implemented with the introduction of targeted biologics and immunotherapeutic agents (e.g., bevacizumab, cetuximab, panitumumab, pembrolizumab), thus expanding the proportion of patients eligible for curative-intent surgery, but their use may lead to side effects or complications[120]. Prediction of tumor response before starting chemotherapy would allow to choose the best treatment, avoiding unnecessary adverse effects of the therapy. Radiomics of CLRM has been proven promising for predicting response to different chemotherapy regimens, but the predictive value of radiomics features seems to be treatment dependent[95]. As shown by two systematic reviews[95,121], most studies performed radiomics on CT rather than MRI. In patients treated with FOLFOX or FOLFIRI, low skewness and narrower standard deviation-both suggesting increased tumor homogeneity-were associated with a high response rate to chemotherapy[122]. Giannini et al[41] developed a radiomics signature to predict behaviour of individual CLRMs to targeted treatment in patients with HER2 amplification undergoing dual target therapy; in their model, the two most important radiomics features were difference in variance and homogeneity, thus again highlighting the role of texture analysis homogeneity of CLRM for assessing lesion diagnosis and outcome. Other studies investigated the role of radiomics in assessing tumor response after other targeted biologics and immunotherapeutic agents, such as regorafenib[117], and bevacizumab[116,48].

CONCLUSION

Nowadays much evidence has revealed that not all clinical risk features are equal, not all affect overall survival, and the decision to treat colorectal cancer with adjuvant chemotherapy should be assessed in multidisciplinary approach[123]. In this scenario, radiomics could play a pivotal role in colorectal cancer workup as an additional tool in clinical setting with the expectancy to help clinicians in identifying patients with high-risk disease. In particular, the main fields examined were the preoperative assessment of the differentiation between low- and high-grade colorectal cancer, and the prediction of nodal metastases[124-131]. The results achieved good and consistent efficiency in identifying high-risk clinical factors, reinforcing the idea that radiomics could play a central role in colorectal cancer patient workup. Nevertheless, radiomics has numerous shortcomings that make daily use extremely difficult. Among these, the lack of standardization and validation, poor reproducibility, and missing prospective multicentric studies represent the main drawbacks that must be overcome to introduce the radiomics approach to the clinical routine[7]. Further studies are needed to assess the performance of radiomics in stratifying patients with high-risk disease in patients with non-metastatic colorectal cancer who could benefit from adjuvant chemotherapy.