Interpretable Artificial Intelligence in Assisting Treatment Response Prediction for Locally Advanced Rectal Cancer...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Preoperative assessment of pathologic complete response (pCR) is crucial for anorectal preservation in locally advanced rectal cancer (LARC).
• Current AI assistance for pCR prediction faces challenges in prospective validation and interpretability.

Purpose of the Study:

• To develop and validate an interpretable AI model (RAPIDS-II) for preoperative pCR assessment in LARC patients.
• To evaluate the performance of RAPIDS-II in predicting pCR and its impact on clinical decision-making.

Main Methods:

• A Deep Residual Shrinkage Network (DRSN) was trained on radiomic features from MRI scans.
• A multimodality model, RAPIDS-II, integrated DRSN's Radscore with clinicopathological factors.
• Model performance was validated retrospectively, in a testing set, and prospectively in a multicenter trial.

Main Results:

• RAPIDS-II demonstrated robust pCR prediction performance with an AUC of 0.795 in the prospective validation cohort.
• The AI tool significantly improved radiologists' visual assessment accuracy, particularly for junior clinicians.
• SHapley Additive exPlanations confirmed Radscore as the primary contributor to RAPIDS-II predictions.

Conclusions:

• The interpretable RAPIDS-II model shows strong performance in pCR evaluation for LARC.
• RAPIDS-II has the potential to assist clinicians in tailoring individualized neoadjuvant therapy strategies.
• The AI tool is particularly beneficial for less experienced radiologists, enhancing treatment planning.