(An) interpretation study on multi-modal network for vertebral compression fracture prediction after stereotactic body radiotherapy

Abstract

A number of studies have predicted side effects after radiotherapy based on statistical analysis. In this work, Vertebral Compression Fracture (VCF) was predicted using a deep learning (DL) network after Stereotactic Body Radiation Therapy (SBRT). Comparisons with various machine learning techniques have verified that deep learning is an appropriate approach. Using LIME and Grad-CAM, the study analyzes how reasonable the prediction process of the DL network is. As a result, the prediction of multi-modal networks utilizing both CT and planned dose maps of patients turned out to be the most reliable. This confirms that it is reasonable to use DL for prediction, and it is expected that VCF can be predicted in the early step and better treatment plans could be implemented.