Paper-Conference

This study presents a fully automated csPCa grade classification framework using bp-MRI for lesion segmentation and radiomics feature extraction. A total of 124 radiomics features were analyzed, with statistically significant features selected to train five machine learning classifiers (SVM, KNN, Bayesian, XGBoost, and Random Forest). The XGBoost classifier achieved the highest accuracy of 96% in distinguishing significant (ISUP 4+5) from non-significant (ISUP 2+3) PCa based on segmentation-predicted maps.

The paper introduces a multiple decoder U-net architecture for improved Organ at Risk (OAR) segmentation in CT images, crucial for radiation treatment planning. It uses segmentation disagreement between decoders as an attention mechanism and incorporates a CT intensity integrated regularization loss to handle low contrast tissues. Tested on two public OAR challenge datasets, the proposed methods significantly enhance segmentation accuracy.