Multimodal Fusion Prediction of Radiation Pneumonitis via Key Pre-Radiotherapy Imaging Feature Selection Based on...

Area of Science:

• Oncology
• Radiotherapy
• Medical Imaging

Background:

• Radiation pneumonitis (RP) is a severe complication for locally advanced non-small cell lung cancer (LA-NSCLC) patients post-thoracic radiotherapy.
• Predicting RP is difficult due to complex risk factors, incomplete data, and lack of slice-level annotations in pre-radiotherapy CT scans.

Purpose of the Study:

• To develop and validate a multimodal fusion framework for accurate RP prediction in LA-NSCLC patients.
• To overcome limitations of unimodal prediction and identify key features in pre-radiotherapy CT images.

Main Methods:

• A retrospective study of 670 LA-NSCLC patients (535 training, 135 testing) who received thoracic radiotherapy.
• Feature extraction using a fine-tuned C3D network for CT images.
• Implementation of a Dual-Layer Attention-Based Adaptive Bag Embedding Multiple-Instance Learning (DAAE-MIL) module for instance screening and representation.
• Multimodal model integration combining clinical data, radiomics, and CT-derived deep features.

Main Results:

• The proposed DAAE-MIL framework achieved high RP prediction performance.
• The multimodal model demonstrated superior accuracy (ACC: 0.93) and AUC (0.97) on the test set compared to state-of-the-art and unimodal methods.

Conclusions:

• The DAAE-MIL multimodal framework effectively predicts RP risk in LA-NSCLC patients.
• This approach addresses single-modal prediction limitations and enhances clinical value for RP risk assessment.