Designing a hybrid optimization methodology for delineating boundary of ultrasound prostate cancer with an explainable mathematical model

Accurately segmenting prostate cancer (PCa) is crucial for enhancing male survival rates, yet it poses challenges due to low-intensity contrast around the PCa outline caused by intestinal gas interference, the existence of shadow artifact, the impact of the heterogeneity among different patients, and human anatomical diversities. Our study proposes a novel ultrasound (US)-guided hybrid optimization algorithm, containing two subnetworks: 1) the first subnetwork uses a deep parallel network structure to complete the initial segmentation stage automatically; 2) the second subnetwork is used to fine-tune the initial outcome, where the initial PCa outlines are optimized via an intelligent hunting polygon tracking method linked to a modified quantum-inspired evolutionary neural network. After the neural network's training, an explainable mathematical mapping formula based on the optimal parameters of an evolutionary neural network is adopted to produce smooth PCa outlines. Experimental outcomes prove the superiority of our proposed methodology over other recent medical image segmentation approaches, achieving an average Dice score (DS), Jaccard index (JI), and accuracy (ACC) of 83.3+ 2.5 %, 82.6+ 3.2 %, and 83.2+ 2.7 %, respectively.

Duke Scholars

Author Jing Cai Radiation Oncology