Staging

Predicting treatment response

Surgery, nCRT, or adjuvant chemotherapy are considered the main therapeutical options for patients affected by LARC. As explained above, HR-MRI represents the most valuable tool for tumour assessment for staging, but it is also fundamental for restaging after therapy [6]. However, it did not show great accuracy in discerning fibrotic scars induced by treatments from minimal residual disease [58].

In this setting, the progress of radiomics and AI in diagnostic imaging has paved the way for the development of new tools to assess the treatment response of RC patients [59–61].

Shin et al. [37] created different radiomics models based on T2WI and DWI to predict pCR in a group of 898 LARC patients who underwent nCRT. They used surgical histopathologic analysis as the reference standard for pCR, which was defined as evaluating only the primary tumour. The authors built three models: T2WI, ADC, and a merged one. Among the three models, the T2WI radiomics one showed higher classification performance with an AUC of 0.82 and sensitivity of 80.0% and was superior to experienced radiologists’ performance in diagnosing pCR [37].

Pang et al. [62] projected a system combining radiomics analysis and DL. The model was based on a post-nCRT T2WI MRI, automatically segmenting an ROI on a “suspicious region”, defined as an area with a distinct possibility of containing a tumour or fibrosis as assessed by radiologists. Their method achieved an AUC of 0.815 on the external validation dataset [62].

Jayaprakasam et al. [38] designed a retrospective study to explore the potential of T2WI MRI radiomics feature analysis of the mesorectal fat to predict the nCRT response in 236 patients with LARC. They created a model that could predict pCR with an AUC of 0.89, a sensitivity of 78%, and a specificity of 85.1% [38].

Interestingly, Feng et al. [39] projected the RAPIDS, an ML model based on three feature sets associated with pCR: radiomics MRI features, pathomics nucleus features, and pathomics microenvironment characteristics. Their objective was to predict pCR in patients with LARC starting from a baseline MRI and whole slide images of H&E-stained biopsy slides. Patients had undergone a pretreatment MRI and nCRT followed by surgery. The accuracy of RAPIDS in predicting pCR in LARC was externally validated in two different cohorts with an accuracy of 0.86 and 0.87, respectively. Finally, RAPIDS showed an AUC of 0.812, a sensitivity of 0.888, and a specificity of 0.74 in predicting pCR in the prospective validation study. These results indicated that RAPIDS could represent a novel tool to assist physicians in the tailored management of patients with LARC [39].

Additionally, Aker et al. [63], Yang et al. [64] also evaluated the applicability of texture analysis to multiple MRI sequences to identify potentially significant imaging biomarkers that can accurately detect patients with pCR, highlighting promising results.

Furthermore, some authors also explored the possibility of developing radiomics nomogram to predict pCR [65, 66].

In particular, Wang et al. [65] created an MRI-based radiomics signature to distinguish good responders and poor responders to nCRT and merged it within a nomogram with MRI T stage, ADC values, and circumferential resection margin. Their nomogram could predict a good response to nCRT with a sensitivity of 71% and a specificity of 88% [65]. Similarly, Liu et al. [66] developed a radiomics nomogram based on the combination of a radiomic signature, pre- and post-treatment MRI (T2WI and DWI) images, and the post-treatment tumour length that managed to reach an accuracy of 94% in predicting pCR.

In general, the results of the different models are promising, even if it is not easy to draw a conclusion from the available evidence. However, an early and accurate prediction of the treatment response could significantly improve the management of patients with LARC and favour the development of tailored treatment.

Genotyping

Radiogenomics represents a recently developed field of imaging whose objective is to obtain genotypic characteristics of a disease through diagnostic images and could optimistically represent a game changer in a radiology-pathology correlation [67].

MSI and KRAS/NRAS/BRAF mutations represent critical, independent prognostic factors in RC patients, and their presence is routinely searched through a genetic test on samples from biopsy or surgery to determine personalised treatment and prognosis [6, 40, 68].

However, biopsy and surgery are invasive, time-consuming techniques with potential complications and depend on specific equipment [40, 41].

Therefore, radiogenomics could represent a non-invasive, low-cost, and time-sparing field of diagnostic imaging that could identify RC genotypic characteristics.

KRAS/NRAS/BRAF are critical proteins in the epidermal growth factor receptor (EGFR) signaling pathway. They control the proliferation, differentiation, and invasion of tumoral cells [69]. Their mutations are responsible for poor response to biological therapy and are linked with a higher risk of developing DM [70, 71]. For this reason, Zhang et al. [72] evaluated the possibility of identifying a radiomics signature from T2WI MRI capable of predicting KRAS status in patients with LARC. The LASSO regression was used to evaluate the associations between the features and gene status. Of the 253 features obtained from T2 images of 83 patients, one feature named X. LL_scaled_std was selected and presented a radiomics-based C-index value of 0.703, suggesting that radiomics features could differentiate KRAS status in these patients [72].

In another paper, Cui et al. [73] used an SVM classifier with T2WI-based radiomics features to evaluate KRAS status in 213 RC patients, providing internal and external validation with an AUC of 0.682 and 0.714, respectively. This result underlines that their radiomics signature could be helpful in predicting KRAS status and may support genomic analysis to establish KRAS expression [73]. Finally, Oh et al. [74] used T2WI MRI-based texture analysis to select three radiomics features significantly associated with KRAS mutational status (P < 0.05) in patients with RC. The three features were used to create a model with a decision tree to evaluate the presence of KRAS mutation. The model comprised four terminal nodes, two of which were able to identify KRAS mutation with a sensitivity, specificity, and accuracy of 84%, 80%, and 81.7%, respectively [74].

MSI is a genetic anomaly subsequent to damaged one or more mismatch repair (MMR) proteins [75].

In a recent paper, Zhang et al. [40] explored the feasibility of using an MRI-based DL model, named 3D MobileNetV2 model, to predict MSI status. The group created three models: one based exclusively on clinical factors, one MRI-based, and another combining clinical and imaging characteristics. The imaging-based and the combined model correctly classified 75.0% and 85.4% of MSI status in the test set, with AUC values of 0.820 and 0.868, respectively. Their T2WI MRI-based model showed a good predictive performance for MSI status in RC patients and may represent a helpful tool to select patients who would benefit from chemotherapy or immunotherapy [40].

Furthermore, Li et al. [76] developed a radiomic model based on MRI images to predict preoperative MSI status in RC patients. They created three main models: a T2-based, an ADC-based, and a combined model, which showed an AUC of 0.895, 0.796, and 0.926 in the testing set, respectively [76].

Assessing high-risk histopathological features

RC may present numerous histopathological features, such as differentiation degree, EMVI, TB, LVI, and PNI that are linked with poor prognosis and have to be taken into account for stratifying the risk of RC patients [5]. A reliable pre-treatment assessment of these factors would probably favour and accelerate the progress of tailored treatment strategies and consequently, the transition toward precision medicine [5, 77–79].

Tumour grading measures cell anaplasia in the sampled tumour and has an essential role in RC prognosis. Well-differentiated tumours generally show better outcomes in this setting [80]. In a recent paper, Meng et al. [81] evaluated the performance of three ML classifiers to recognise well-differentiated RCs on the basis of radiomics characteristics obtained from T2WI, DWI, and DCE MRI sequences. The LASSO algorithm outperformed the other two ML classifiers, with an AUC of 0.72 in the validation group [81]. Furthermore, He et al. [82] developed a feature selection method and ML-based prediction model using MRI-based radiomics features to discern among different pathological grades for RC. The model showed relatively acceptable performance in tumour grading of RC with AUC values of grades 1, 2, 3, and 4 of 0.717, 0.683, 0.690, and 0.827, respectively, in the validation set [82].

TB is an emerging prognostic biomarker in RC. TBs are defined as single cancer cell or clusters of up to four cancer cells located at the invasive tumour front [83]. Recently a group of researchers constructed a model based on multiple MRI-sequences radiomics to identify TB in LARC before surgery, which presented an accuracy of 81.2% [42].

MRI is also considered one the most valuable tool to establish the presence of EMVI; however, HR-MRI performances in evaluating EMVI are still not that high [84].

For this reason, Zhao et al. [85] built a radiomic nomogram based on multiple MRI sequences (T1WI, T2WI, and proton density) to identify EMVI, with an AUC of 0.899, which resulted in being superior to radiologists.

Another interesting nomogram was created by Yu et al. [86], including a DCE MRI radiomics signature and clinical features. They found that the nomogram performed better than quantitative perfusion parameters such as Ktrans in predicting EMVI, with 88.9% sensitivity and 78.3% specificity in the test dataset [86].

Moreover, Shu et al. [87] projected a system based on T1WI, T2WI, DWI, and DCE MRI sequences to detect EMVI. The algorithm was also merged with clinical features and presented an AUC of 0.835, a sensitivity of 0.714, and a specificity of 0.885 [87].

LVI and PNI rapresent two negative prognostic factors in RC. LVI and PNI occur when cancer invades the layers of the lymphovascular wall or spreads along the nerve sheath [88].

Zhang et al. [89] designed a multi-modality radiomic nomogram based on T2WI and DWI MRI sequences plus enhanced CT images to predict the presence of LVI in RC patients. Their nomogram had significant predictive power in the validation cohort with an AUC of 0.876. Instead, Guo and his group [90] operated similarly to project radiomic nomograms to predict PNI in RC patients, obtaining an AUC of 0.884 in the test dataset.

Finally, Chen et al. [91] realised a different radiomics nomogram based only on oblique T2WI MRI sequences that could predict PNI status with an accuracy of 0.71 and an AUC of 0.85.

Prognosis

Nowadays, although the progress of surgery and CRT protocols have considerably improved the outcomes of patients affected by RC, local recurrence (LR) and DM still represents two critical negative prognostic factors [92, 93]. HR-MRI has a pivotal role in detecting LR during follow-up scans in this setting. Recently, researchers tried to understand the potential of HR-MRI-based radiomics and AI to predict LR in patients with RC [38]. Jayaprakasam et al. [38] developed a T2WI MRI radiomics-based model to predict LR in 236 patients who underwent nCRT. The eight most significative radiomics features were extracted from axial T2WI of mesorectal fat to develop a model using SVM. They found that it could predict LR with a sensitivity and specificity of 68.3% and 80.7%, respectively [38].

DM represents another critical prognostic factor that influences the outcomes of RC patients. Recently, many authors developed different promising radiomic signatures from segmented tumour regions based on multiple HR-MRI sequences that could be used to predict DM-free survival (DMFS) in patients with LARC [94–96].

Moreover, a group of researchers recently designed a retrospective study to explore the capability of an AI radiomics-based model to predict 3-year DMFS in patients with LARC after receiving nCRT, with a C-index of 0.747 and an AUC of 0.894 [43]. Another group built up a radiomic nomogram constructed on an MRI-based radiomic signature that exhibited high performance in predicting DMFS, with C-indices of 0.848, 0.831, and 0.825 in the test groups, respectively [97]. Furthermore, Liu et al. [36] investigated the possibility of predicting SDM in RC patients via a clinical-radiomics model before surgery. They realised a model based on the preoperative T2WI MRI images of 177 patients that showed an AUC of 0.827, which proved that the model could help stratify patients with a higher risk of SDM before surgery [36].