Purpose: Microultrasound might serve as a viable alternative to MRI for initial prostate cancer detection. This study developed a deep learning model to detect clinically significant prostate cancer (csPCa) on microultrasound and evaluated its impact on diagnostic performance among novice readers. Materials and Methods: One hundred microultrasound sweeps (50 csPCa, 50 benign) were used to train a deep learning model. Model performance was assessed using the area under the receiver operating characteristics curve (AUC) and Dice similarity coefficient (DSC). In an independent test set (25 csPCa, 25 benign), 5 novice readers each evaluated 2 unique sets of 10 sweeps before and after a structured microultrasound training. In both sessions, they first reviewed sweeps unaided and then with algorithm-generated heatmaps. Results: On the training set, DeepTRUS achieved an AUC of 0.687 (95% confidence interval 0.583-0.791) and a DSC of 0.30 (0.21-0.39). In the test set, reader AUC improved from 0.459 (0.294-0.624) to 0.556 (0.394-0.718) before training, and from 0.610 (0.452-0.767) to 0.782 (0.656-0.908) after training, with DeepTRUS assistance. Sensitivity increased from 60% to 72% pretraining and from 80% to 88% posttraining. The interaction between training and DeepTRUS assistance was significant ( P = .006). Conclusions: This pilot study demonstrates the feasibility of training deep learning models on microultrasound data and their potential to enhance diagnostic performance when combined with reader training. The findings provide proof of concept for artificial intelligence–assisted microultrasound interpretation and inform the design of larger validation studies.
Haney-Aubert et al. (Wed,) studied this question.