JMIR Publications

ABSTRACT

Background:

Rectal cancer (RC) is a prevalent malignancy with varying responses to radiotherapy (RT), highlighting the need for accurate prediction tools to optimize treatment plans.

Objective:

To develop a deep learning-based radiomics model that predicts the response of RC patients to RT, potentially enhancing personalized treatment strategies.

Methods:

We collected radiomics imaging data from 2000 RC patients, encompassing CT and MRI scans. The data underwent preprocessing, including denoising, normalization, and segmentation via a U-Net model. We evaluated ten deep learning algorithms, including CNNs, ResNets, and Transformer models, focusing on the Transformer model's self-attention mechanism for feature extraction and prediction accuracy. The dataset was divided into training, validation, and test sets (8:1:1 ratio), with k-fold cross-validation ensuring model reliability.

Results:

The Transformer model achieved the highest accuracy of 87% in predicting RT response, outperforming CNNs (82%) and GCNs (84%). Its performance was validated using ROC curves, with an AUC of 0.93. Grad-CAM visualization confirmed the model's focus on critical tumor regions, enhancing interpretability and clinical relevance.

Conclusions:

The deep learning-based radiomics model, especially the Transformer model, shows significant potential in predicting RC patients' responses to RT. This study provides a robust method for predicting RT response, contributing to personalized treatment plans and improving clinical decision-making. Future research should focus on expanding the dataset and integrating multimodal data to further enhance model performance and applicability. Clinical Trial: Not applicable.