OBJECTIVES: The use of prostate magnetic resonance imaging (MRI) is increasing, and coverage often captures substantial portions of the pelvis, visualizing findings outside the prostate gland. The objective of this study was to evaluate the feasibility of automated detection and segmentation of high-prevalence incidental findings in prostate MRI. MATERIALS AND METHODS: This IRB-approved, retrospective study included n=465 prostate MRI examinations (1.5 and 3.0 T), comprising n=315 internal cases from our institution and n=150 external cases from 3 independent data sets. Manual segmentations were performed for perirectal lymph nodes, sigmoid diverticulosis, urinary bladder diverticula, bladder wall thickenings, inguinal hernias, periarticular bone changes of the hip, prominent synovial compartments of the hip, and hydroceles testis on axial T2-weighted (T2w) images for n=265 internal cases (n=520 ROIs). An nnU-Net model was trained on n=213 of these cases. The remaining n=52 independent cases (quantitative test set) were used for the quantitative evaluation of model performance using Dice score, intersection over union (IoU), Hausdorff distance (HD), mean surface distance (MSD), confusion matrices, sensitivity, specificity, and accuracy. Furthermore, n=200 additional examinations (reader test set), comprising n=50 internal, n=150 from 3 independent external data sets, were evaluated by 2 radiologists in an AI-assisted evaluation. Readers assessed the presence of incidental findings (step 1) and the correctness of the nnU-Net-predicted findings (step 2), based on the AI-predicted segmentations. RESULTS: Segmentation performance varied between the incidental findings. Evaluation of the independent quantitative test set revealed that the highest mean Dice scores were achieved for sigmoid diverticulosis (0.80±0.14), hydroceles testis (0.76±0.20), and periarticular bone changes of the hip (0.70±0.07). Radiologists' evaluation of AI predictions on an independent reader test set comprising 1 internal and 3 external data sets demonstrated high agreement in AI-assisted evaluation for most incidental findings. Accuracies per data set (PROSTATEx/Ai4ar/internal/Prostate-3T) were 0.94/0.62/0.96/0.94 for perirectal lymph nodes, 0.82/0.80/0.80/0.84 for sigmoid diverticulosis, 0.86/0.86/0.98/0.98 for urinary bladder diverticula, 0.96/0.82/0.88/0.72 for bladder wall thickenings, 0.94/0.82/0.78/0.86 for inguinal hernias, 0.90/0.80/0.82/0.84 for periarticular bone changes of the hip, 0.90/0.82/0.84/0.96 for prominent synovial compartments of the hip, 0.98/0.80/0.96/0.96 for hydroceles testis. Inter-reader agreement for the AI-assisted evaluation of the presence of incidental findings on T2w images was high, with Cohen κ values ranging from 0.74 to 0.92 for most findings. CONCLUSIONS: The nnU-Net-based AI model was able to capture and segment frequent incidental findings in prostate MRI across 4 independent data sets, demonstrating potential to support radiologists in consistent reporting. This supports further research with larger, more diverse data sets, including additional annotations and clinical targets.
Skwierawska et al. (Fri,) studied this question.