Multi-Modal Siamese Network for Diagnostically Similar Lesion Retrieval in Prostate MRI

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Multi-parametric prostate MRI (mpMRI) aids prostate cancer diagnosis but is challenging for radiologists.
• Comparing current MRIs with similar diagnosed cases is a standard but time-consuming procedure.
• Computerized Content-Based Image Retrieval (CBIR) systems can enhance accuracy and efficiency in radiological reporting.

Purpose of the Study:

• To introduce a novel supervised siamese deep learning architecture for prostate cancer diagnosis.
• To enable the system to process multi-modal and multi-view MR images with similar PIRADS scores.
• To improve the accuracy and workflow of radiological image interpretation.

Main Methods:

• Development of a supervised siamese deep learning architecture for CBIR.
• Handling of multi-modal and multi-view MR images.
• Experimental comparison against standard siamese networks and autoencoders.

Main Results:

• The proposed multi-view siamese network demonstrated significantly improved diagnostic performance (ROC-AUC).
• Enhanced information retrieval metrics including Precision-Recall, DCG, and mAP were achieved.
• The system outperformed established deep learning-based CBIR methods.

Conclusions:

• The novel multi-view siamese network offers superior performance for retrieving similar prostate MRI cases.
• This approach enhances diagnostic accuracy and reporting efficiency for prostate cancer.
• The architecture's general design supports broad applications in diagnostic medical imaging retrieval.