A polyline searching-driven evolutionary AI for disease detection of medical imaging data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate ultrasound (US) prostate cancer (PCa) segmentation is crucial for interventional guidance and diagnosis.
• Challenges include blurred boundaries, US shadow artifacts, and patient-specific morphological variations.
• Existing methods struggle with the complexities of US PCa imaging.

Purpose of the Study:

• To develop a novel coarse-to-fine optimization framework for enhanced US PCa segmentation.
• To address limitations of current segmentation techniques in US imaging.
• To improve diagnostic accuracy and treatment efficacy for prostate cancer.

Main Methods:

• A four-stage framework: deep learning for region of interest (ROI) localization, principal curve-based polyline search for vertex sequence determination.
• Enhanced quantum evolutionary algorithm for optimal backpropagation neural network initialization.
• Mathematical model defining ROI boundary using neural network parameters.

Main Results:

• Achieved superior performance on US PCa data with mean Dice similarity coefficient (DSC) of 83.6±3.1%, Jaccard (OMG) of 71.8±2.5%, and accuracy (ACC) of 83.5±3.1%.
• Demonstrated best segmentation capability on the public Digital Database of Thyroid Ultrasound Images (DDTI) dataset.
• Outperformed current algorithms in segmentation accuracy and robustness.

Conclusions:

• The proposed intelligent algorithm offers a significant advancement in US PCa segmentation.
• Accurate segmentation aids in protecting vulnerable anatomical structures during clinical procedures.
• The method is expected to enhance diagnostic accuracy and treatment outcomes for prostate cancer.