Editorial for “Selecting candidates for organ-preserving strategies after neoadjuvant chemoradiotherapy for rectal cancer: development and validation of a model integrating MRI radiomics and pathomics”

A key clinical question in rectal cancers today is identification of patients with pathologic good response after neoadjuvant chemoradiotherapy, towards enabling “watch-and-wait” or organ preserving management. Non-operative management of rectal cancer patients achieving excellent response and tumor downstaging after neoadjuvant therapy could allow them to avoid surgical morbidities associated with a total mesorectal excision while not impacting their recurrence or survival rates (1).

Leveraging the fact that MRI forms a key component in identifying such patients (2), there has been a significant body of work demonstrating the utility of radiomics (i.e., extraction of quantitative measurements from radiographic imaging) in identifying rectal cancer patients who have achieved pathologic complete response (pCR), using pre- or post-treatment MRI (3, 4). Initial results have been presented on quantitative measurements from digital pathology biopsy specimens (known as pathomics) being associated with pCR in rectal cancers (5).

In this issue of JMRI, Wan, Sun, and colleagues (6) have presented a machine learning model incorporating radiomic, pathomic, and clinicoradiological variables towards identifying pCR after neoadjuvant chemoradiotherapy in patients with locally advanced rectal cancer. Using a total of 153 patients split across discovery and validation cohorts, radiomic features were extracted from pre- and post-treatment T2-weighted and diffusion-weighted (DWI) MRI (using the Radcloud platform). Pathomic features were extracted from automatically identified tumor regions on pre-treatment biopsy images as responses to a pretrained VGG19 deep learning network. In addition, clinicoradiological factors such as clinical tumor/nodal stage and tumor dimensions were also evaluated. Machine learning analysis revealed that combining clinicoradiologic parameters, pre- and post-treatment radiomics, as well as pathomics yielded the best overall AUC in identifying pCR after neoadjuvant chemoradiotherapy; as compared any alternative feature combination. Individually, pre- and post-treatment radiomic features yielded relatively similar performance for predicting pCR while biopsy pathomic features yielded a marginally lower AUC in this task. Decision and calibration curve analysis were used to further validate the performance of the multiscale model in predicting pCR. These findings and overall performance metrics align with recent large-scale studies on predicting pCR and survival in rectal cancers using pre-treatment radiomic and biopsy-based pathomic features (7, 8).

Critical to developing such predictors is interrogating which image features and patterns were identified as relevant in making a classification of pCR. Of interest here is that clinically used radiologist assessments of tumor thickness and length on pre- and post-treatment MRI were not only significantly associated with pCR but also augmented predictor performance (albeit marginally). This suggests complementarity of expert evaluation to computerized approaches, which could have implications for wider clinical adoption of radiomic approaches. A majority of radiomic features identified here involved gray level co-occurrence (GLCM) and size zone matrix (GLSZM) responses, which capture heterogeneity within local neighborhoods of the tumor region on MRI (aligning with features reported in previous investigations (3, 4)). The authors similarly interrogated the bases of selected pathomic features (visualized in Supplementary Material) as capturing responses within different parts of the annotated tumor on biopsy specimens. However, leveraging of responses from a pre-trained VGG19 network may be considered relatively less intuitive in comparison to specific pathomic measurements of nuclear shape/texture or proportions of tissue types within the biopsy microenvironment, as have been utilized in previous studies (7, 8). A potential avenue to expand on the current study may also be to leverage responses from the specialized DeepLabv3+ segmentation model that was used to accurately identify tumor regions on biopsy images in this work. This in turn could allow us to determine biological associations across scales in more detail (9, 10), thus identifying explainable relationships between radiomic and pathomic features towards more comprehensive characterization of rectal cancer.

For the next step of determining clinical utility, it will be important to validate these findings on larger, multi-institutional cohorts. This would also allow for robust interrogation of how imaging parameters, field strength, and biopsy slide preparation (among other factors) impact the generalizability and reproducibility of such single institution studies. The findings in this study contribute to the growing understanding of pCR and tumor response in rectal cancers and motivate the potential for building integrated radiomic, pathomic, and genomic predictors in the future.