Dosiomics-guided deep learning for radiation esophagitis prediction in lung cancer: optimal region of interest definition via multi-branch fusion auxiliary learning.

Abstract

Accurate delineation of regions of interest (ROIs) is critical for feature extraction and selection in radiomics-based prediction models. To develop a combined dosiomics and deep learning (DL) model for predicting grade ≥ 2 radiation esophagitis (RE) in lung cancer patients undergoing radiotherapy, we propose a multi-task auxiliary learning approach to define accurate and objective ROIs based on radiation dose distribution (RDD) images. Lung cancer patients who underwent radiotherapy were gathered retrospectively from hospital 1 (January 2020 and December 2022) for model development. Two external validation sets were obtained retrospectively from hospital 2 (January 2021 and December 2022) and hospital 3 (January 2022 and December 2023), respectively. A dosiomics-guided deep learning (DGD) network using multi-task auxiliary learning to define accurate and objective ROIs was introduced by integrating dosiomic features with high-dimensional DL features for RE prediction. This study enrolled 488 patients from three hospitals: 235 in the training set, 101 in the internal validation set, 57 in the external validation set 1 and 95 in the external validation set 2, respectively. The dosiomics -guided ResNet34 combined with contrastive learning and auxiliary segmentation module achieved the best AUCs of 0.88 [95% CI: 0.76-0.95], 0.82 [95% CI: 0.65-0.96], 0.83 [95% CI: 0.74-0.92] in the internal validation set, external validation set 1, and external validation set 2, respectively. The proposed DGD model leverages multi-task auxiliary learning to automatically define ROIs and effectively predict RE in lung cancer patients undergoing radiotherapy.