Treatment decision support for esophageal cancer based on PET/CT data using deep learning

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Precise treatment decisions are crucial for esophageal cancer patients.
• Traditional methods struggle with complex PET/CT imaging patterns.
• Deep learning offers advanced predictive modeling capabilities.

Purpose of the Study:

• To develop a novel deep learning model for esophageal cancer treatment decision support.
• To integrate convolutional and transformer components for enhanced PET/CT data analysis.
• To improve predictive accuracy beyond existing models.

Main Methods:

• A hybrid deep learning architecture combining convolutional and transformer elements was proposed.
• Key components include a Convolutional Feature Extractor, Multi-scale Pooling, and a Multilayer Perceptron.
• The model was evaluated using AUCROC, F1 score, and Balanced Accuracy, benchmarked against state-of-the-art models.

Main Results:

• The proposed model achieved superior performance with an AUCROC of 0.9935 and Balanced Accuracy of 0.9630.
• Ablation studies confirmed the necessity and effectiveness of each architectural component.
• The hybrid approach demonstrated superior predictive accuracy over existing methods.

Conclusions:

• A novel hybrid deep learning architecture was developed for esophageal cancer treatment decision support.
• The model leverages multi-scale spatial encoding for enhanced predictive accuracy.
• Tailored architectural innovations significantly improve upon existing methods.