Improve Model Performance Rectal Cancer Staging by Radiomics Analysis of Diffusion-weighted Imaging and T2-weighted Imaging

Motivation: Accurate preoperative tumor staging is important for treatment decisions of rectal cancer (RC). Radiomics can be helpful in identifying tumor stages. Goal(s): To improve model performance of tumor staging (≥T3 or <T3) in RC patients by radiomics. Approach: Five image contrasts, including diffusion-weighted, apparent diffusion coefficient, T2-weighted, and their combinations, were analyzed. The targeting regions were determined by five strategies: bounding box, manually determined tumor region of interest (ROI), and ROIs with dilation of 1, 3, and 5 pixels. Results: The DWI_b800+T2WI and DWI+FS-T2WI with bounding box and CatBoost achieved an accurate AUC=0.889 for identifying RC patients ≥T3. Impact: Compared to extracting features from DWI or T2WI solely, radiomics analysis of combinations of DWI_b800+T2WI or DWI_b800+FS-T2WI exhibited better AUC and TNR in identifying rectal cancer patients ≥T3. A bounding-box was the best strategy for determining the analyzed region.