# Predicting treatment response to neoadjuvant chemotherapy in locally advanced rectal cancer: A combined deep learning and machine learning approach utilizing longitudinal multi-sequence MRI

## TL;DR

This study uses deep learning and MRI scans to predict how well rectal cancer patients will respond to chemotherapy before surgery.

## Contribution

First study to use longitudinal multi-sequence MRI and deep learning to predict treatment response in rectal cancer.

## Key findings

- A fusion deep learning model achieved an AUC of 0.846 for predicting treatment response.
- Post-chemotherapy DWI deep learning signatures were the strongest predictors of favorable response.
- Adding clinical features slightly improved performance but not significantly.

## Abstract

To develop and validate deep leaning-based machine learning models using longitudinal multi-sequence MRI for predicting treatment response of patients with locally advanced rectal cancer (LARC) to neoadjuvant chemotherapy (NCT).

This retrospective study included 169 LARC patients who received 2–4 cycles of CAPOX chemotherapy before surgery. Patients were randomly divided into a training cohort (n = 118) and a test cohort (n = 51). High-resolution paired MRI sequences (CE-T1WI, T2WI, DWI) were acquired before and after NCT. These sequences were then independently input into a DenseNet121 network to generate predictive probability scores, which served as deep leaning (DL) signatures. Prediction models were built using an SVM classifier. These models were built either by integrating the deep learning signatures alone or by combining them with clinical and radiological features. Model performance was assessed using AUC, accuracy, sensitivity, and specificity. Calibration was evaluated with calibration plots and Brier scores, and clinical utility was analyzed via decision curve analysis (DCA).

In the test cohort, the fusion DL model, integrating pre- and post-NCT multi-sequence DL signatures, achieved an AUC of 0.846. The combined clinical-radiological-deep learning (CRD) model, which added clinical-radiological features to the fusion DL model, reached the highest AUC of 0.851, but the improvement was not statistically significant.

The fusion DL model showed strong performance in predicting pathological response in LARC. The post_DWI signature was the main contributor to the model.

•First study to predict LARC response after neoadjuvant chemotherapy using longitudinal multi-sequence MRI deep learning.•Longitudinal fusion of pre-&-post multi-sequence MRI lifts AUC to 0.846 without extra clinical data.•SHAP analysis confirms post-NCT DWI DL-signature as the strongest biomarker of favorable pathological response.•Adding post-CRM & maximum wall thickness yields peak AUC 0.851, but gain is not statistically significant.

First study to predict LARC response after neoadjuvant chemotherapy using longitudinal multi-sequence MRI deep learning.

Longitudinal fusion of pre-&-post multi-sequence MRI lifts AUC to 0.846 without extra clinical data.

SHAP analysis confirms post-NCT DWI DL-signature as the strongest biomarker of favorable pathological response.

Adding post-CRM & maximum wall thickness yields peak AUC 0.851, but gain is not statistically significant.