Alberto Rossi (DIISM, University of Siena)
Jul 31, 2019 – 11:00 AM
DIISM, Artificial Intelligence laboratory (room 201), Siena SI
Prostate Multi-parametric MRI (mpMRI) is a great tool to diagnose prostate cancer but is difficult to interpret even for expert radiologists. A common radiological procedure to analyze a difficult case is to compare it to diagnostically similar cases. Computerized Content-Based Image Retrieval system (CBIR) can help improve workflow and increase accuracy. A simple CBIR like Autoencoder (AUTO) is less suitable as it focuses mainly on image appearance and not on diagnostic information and is limited to single image views. New deep learning-based CBIR is represented by Siamese network (SIAM). We developed an enhanced SIAM that can retrieve images with diagnostically similar PIRADS score and extends similarity to multi-modal and multi-view MRI. We compare the enhanced SIAM to baseline SIAM and AUTO and study the effect of supervised diagnostic training and number of views. CBIRs were trained on 890 annotated lesion data set and evaluated the results using diagnostic (ROC-AUC), and information retrieval metrics (Precision Recall – PR, Discounted Cumulative Gain – DCG). The AUC results are 0.47, 0.76, 0.86, 0.83, respectively for AUTO, and SIAM using 1, 2 and 3 views. In conclusion, our enhanced SIAM significantly improves over existing CBIR in mpMRI and reaches a level where it can be an aid for radiologists to interpret mpMRI. Our enhanced SIAM is generic and applicable to other diagnostic medical imaging applications.