Enhancing the Accuracy of Tumor Response Assessment in Rectal Cancer with Radiomics and Deep-Learning: Insights from the SFX Trial

Background: After total neoadjuvant therapy (TNT) for rectal cancer, eligibility for a surgery-sparing watch-and-wait (WW) strategy relies on clinical assessment including endoscopy, digital rectal exams, and MRI. Accurate assessment of tumor response during post-TNT surveillance has not been optimized, leading some patients to unnecessarily undergo proctectomy while others experience tumor regrowth or recurrence. 

Objective: To accurately determine tumor response to TNT by integrating radiomics and deep-learning (DL) in MRI assessment. 

Methods: We partnered with the SFX trial which enrolled 37 patients with stage II/III rectal cancer.  Each patient was treated with TNT with short-course radiation and FOLFOXIRI. Protocolized assessment included MRIs for (i) initial staging of rectal cancer (ii) re-staging post-radiation, and (iii) re-staging post-FOLFOXIRI. Patients with clinical complete response were offered WW; patients with an incomplete clinical response were recommended for proctectomy with total mesorectal excision. We performed imaging pre-processing and 3D manual segmentation of rectal tumors as well as radiomic feature extraction, stability assessment, and selection. 

Results: We extracted 1302 radiomic features, including first-order statistics, shape, texture, and intensity values, from MRIs, which we will analyze longitudinally. Using DL, we are exploring the connections between variations in the radiomic features and patient outcomes. We anticipate that patients who achieve clinical complete response will display consistent radiomic signatures compared to patients who do not. Our integration of radiomics and DL is intended to reduce the inherent subjectivity in the interpretation of MRIs. 

Conclusion: Radiomics and DL will improve diagnostic accuracy after TNT for rectal cancer, offering valuable tools to accurately determine clinical complete response.