Deep learning combining imaging, dose and clinical data for predicting bowel toxicity after pelvic radiotherapy

Area of Science:

• Radiation oncology
• Medical imaging
• Machine learning

Background:

• Predicting radiotherapy toxicity is complex due to the need for multimodal data analysis.
• Integrating 3D imaging and clinical data for toxicity prediction presents significant challenges.

Purpose of the Study:

• To develop a deep learning model for simultaneous analysis of computed tomography (CT) scans, dose distributions, and clinical metadata.
• To predict radiotherapy-induced toxicity and identify key clinical risk factors and critical anatomical regions.

Main Methods:

• A deep model utilizing multiple instance learning with feature-level fusion and attention was developed.
• The model was trained on data from 313 patients treated with 3D conformal radiation therapy and volumetric modulated arc therapy, including CT scans, dose distributions, and clinical data.
• Patient-reported late bowel toxicity data was used for model training.

Main Results:

• The model successfully identified potential risk factors and critical anatomical regions associated with toxicity.
• Joint analysis of clinical, imaging, and dose data improved prediction for bowel urgency (AUC 88%) and fecal incontinence (AUC 78%).
• Diarrhea prediction showed best performance (AUC 68%) using clinical features alone.

Conclusions:

• Feature-level fusion and attention mechanisms enable effective multimodal data analysis in radiotherapy toxicity prediction.
• The model provides interpretability by explaining input contributions and detecting spatial associations with toxicity.