Toxicity Prediction in Pelvic Radiotherapy Using Multiple Instance Learning and Cascaded Attention Layers.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Current radiotherapy planning relies on simplified dose-toxicity assumptions.
• Understanding radiation-induced toxicity, especially patient-reported outcomes, remains a challenge.
• Individualized treatment optimization requires deeper insights into dose-response relationships for normal tissues.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) using multiple instance learning for predicting toxicity in pelvic radiotherapy.
• To investigate the spatial distribution of toxicity by analyzing the association between radiation dose and patient-reported outcomes.
• To enhance the interpretability of AI models in radiotherapy by segregating attention over space and imaging features.

Main Methods:

• A dataset of 315 patients undergoing pelvic radiotherapy was analyzed.
• Included data: 3D dose distributions, pre-treatment CT scans with annotated structures, and patient-reported toxicity scores.
• A novel CNN architecture with independent attention mechanisms for spatial and feature analysis was proposed.

Main Results:

• The proposed network achieved 80% accuracy in predicting patient toxicity.
• Spatial attention analysis revealed significant associations between radiation dose to the anterior and right iliac regions and toxicity.
• The model demonstrated outstanding performance in toxicity prediction, localization, and explanation, with good generalization capabilities.

Conclusions:

• The developed AI model effectively predicts and explains radiotherapy-induced toxicity.
• Identifying specific anatomical regions associated with toxicity can inform treatment planning.
• This approach offers a promising tool for personalized radiotherapy and reducing side effects.