# A pelvis MR transformer-based deep learning model for predicting lung metastases risk in patients with rectal cancer

## TL;DR

This study develops a deep learning model using pelvic MR images and clinical data to predict lung metastases risk in rectal cancer patients.

## Contribution

A transformer-based deep learning model is proposed for predicting rectal cancer lung metastases using pelvic MR images.

## Key findings

- The transformer-based model achieved an AUC of 83.74% in predicting lung metastases.
- For T4 and N2 stage cancers, the model achieved AUCs of 96.67% and 96.83%, respectively.
- The model outperformed other state-of-the-art deep learning methods in predictive accuracy.

## Abstract

Accurate preoperative evaluation of rectal cancer lung metastases (RCLM) is critical for implementing precise medicine. While artificial intelligence (AI) methods have been successful in detecting liver and lymph node metastases using magnetic resonance (MR) images, research on lung metastases is still limited. Utilizing MR images to classify RCLM could potentially reduce ionizing radiation exposure and the costs associated with chest CT in patients without metastases. This study aims to develop and validate a transformer-based deep learning (DL) model based on pelvic MR images, integrated with clinical features, to predict RCLM.

A total of 819 patients with histologically confirmed rectal cancer who underwent preoperative pelvis MRI and carcinoembryonic antigen (CEA) tests were enrolled. Six state-of-the-art DL methods (Resnet18, EfficientNetb0, MobileNet, ShuffleNet, DenseNet, and our transformer-based model) were trained and tested on T2WI and DWI to predict RCLM. The predictive performance was assessed using the receiver operating characteristic (ROC) curve.

Our transformer-based DL model achieved impressive results in the independent test set, with an AUC of 83.74% (95% CI, 72.60%-92.83%), a sensitivity of 80.00%, a specificity of 78.79%, and an accuracy of 79.01%. Specifically, for stage T4 and N2 rectal cancer cases, the model achieved AUCs of 96.67% (95% CI, 87.14%-100%, 93.33% sensitivity, 89.04% specificity, 94.74% accuracy), and 96.83% (95% CI, 88.67%-100%, 100% sensitivity, 83.33% specificity, 88.00% accuracy) respectively, in predicting RCLM. Our DL model showed a better predictive performance than other state-of-the-art DL methods.

The superior performance demonstrates the potential of our work for predicting RCLM, suggesting its potential assistance in personalized treatment and follow-up plans.