The Role of Artificial Neural Networks in Prostate Magnetic Resonance Imaging (MRI) Segmentation

Introduction In recent years, artificial intelligence (AI)-generated prostate whole-gland segmentation has shown promise for clinical use. This study compares AI- and human-derived prostate segmentation on magnetic resonance imaging (MRI) and evaluates its practical application. Methods A retrospective study from December 2020 to December 2022 evaluated 31 randomly selected patients who had previously undergone MRI-ultrasound (US)-guided fusion biopsies. AI-generated auto-contours of the whole-gland prostate, seminal vesicle, and urethra were produced using the ProtégéAI feature of the MIM Software from MRIs obtained on GE Signa HDxt 3.0T and Siemens Altea Magnetom 1.5T scanners. The same MRIs were independently contoured manually by a board-certified urologist (U) and a board-certified radiologist (R) using MIM Software contouring tools. Volumetric conformity among the three contour sets was assessed using the Dice similarity coefficient, Hausdorff distance (HD), and mean distance to agreement (MDA). Contouring time was recorded for each method, with AI timed from command execution to file generation and physicians timed from file opening to save. Pairwise Wilcoxon signed-rank tests (JMP Pro 15) compared contouring times and similarities across AI to urologist (AI-U), AI to radiologist (AI-R), and urologist to radiologist (U-R) contours. Results The average volumetric Dice similarity, HD, and MDA for the AI-U contours were 0.875±0.039, 9.186±4.004 mm, and 1.410±0.511 mm for the whole-gland prostate, 0.283±0.185, 18.117±13.265 mm, and 4.907±4.992 mm for the urethra, and 0.377±0.244, 14.708±9.489 mm, and 4.260±4.092 mm for the seminal vesicle. For the AI-R contours, the values were 0.757±0.072, 17.562±7.240 mm, and 3.050±1.339 mm for the prostate, 0.162±0.105, 19.956±11.962 mm, and 4.798±3.879 mm for the urethra, and 0.451±0.219, 16.069±9.245 mm, and 4.075±3.734 mm for the seminal vesicle. For the U-R contours, the values were 0.769±0.070, 15.109±6.177 mm, and 2.733±1.265 mm for the prostate, 0.144±0.091, 13.560±7.087 mm, and 3.406±1.985 mm for the urethra, and 0.471±0.216, 12.981±6.756 mm, and 3.264±2.308 mm for the seminal vesicle. Using the Wilcoxon signed-rank test with statistical significance defined as p<0.01, the AI-U Dice similarity for the prostate differed significantly from both the AI-R and U-R comparisons. Similarly, AI-U demonstrated significantly different HD values than both AI-R and U-R, whereas the AI-R and U-R HD comparison was statistically insignificant. All pairwise MDA comparisons for the prostate were statistically significant. For the urethra, the AI-U Dice similarity differed from the other two comparisons, while for HD, the AI-R and U-R comparisons differed significantly from each other. No significant differences were observed among the three comparisons for the seminal vesicle across any metric. The average times to produce contours for the AI, urologist, and radiologist were 96.5 seconds, 285.8 seconds, and 217.9 seconds, respectively (p<0.01 for each time comparison). Conclusion This study suggests that AI may be a useful tool for prostate segmentation workflows in streamlining MRI-US prostate cancer diagnostics by producing similar contouring results in less time. Additional investigation should be conducted regarding the differences between the pathological outcomes of AI and non-AI contours, and the accuracy, cost analysis, and efficiency of AI technology should be elucidated.