Presented by: Puranjay Shori BS
UTHealth Houston School of Public Health

Introduction

This study compares artificial intelligence (AI) derived and human derived whole gland segmentation of the prostate, seminal vesicle, and urethra through magnetic resonance imaging (MRI) and evaluates the practical use of this technology to streamline workflow and care through procedures such as MRI-ultrasound fusion prostate biopsies.

Materials

The AI generated “auto contours” for prostate MRIs of 31 patients were generated through a medical image software to produce contours of the prostate, seminal vesicle, and urethra. These MRIs were also contoured manually by both a board-certified urologist and a radiologist, and the volumetric conformity of the 3 groups of contours was evaluated through the the Dice and Jaccard coefficients, as well as other metrics such as the Hausdorff distance (HD), and Mean Distance to Agreement (MDA). The elapsed time to produce contours was also recorded.

Results

,

The average volumetric Dice and Jaccard coefficients, HD, and MDA for the AI and urologist contoured images (AI-U) were 0.875, 0.779, 9.186 mm, and 1.410 mm for the whole gland segmentation of the prostate, 0.377, 0.260, 14.708 mm, and 4.260 mm for the seminal vesicle, and 0.283, 0.178, 18.117 mm, and 4.907 mm for the urethra; for the AI and radiologist contours (AI-R), the values were 0.757, 0.614, 17.562 mm, and 3.050 mm for the prostate, 0.451, 0.315 16.069 mm, and 4.075 mm for the seminal vesicle, and 0.162, 0.092, 19.956 mm, and 4.798 mm for the urethra. In comparing the contours of the urologist to the radiologist (U-R), the values of the prostate were 0.769, 0.630, 15.109 mm, and 2.733 mm, for the seminal vesicle, 0.471, 0.333, 12.981 mm, 3.264 mm, and for the urethra, 0.144, 0.080, 13.560 mm, and 3.406 mm. For the prostate and urethra, the AI-U comparison was different than the other comparisons (p<0.05). For the seminal vesicle, no statistical difference in the 3 Dice coefficients was observed (p>0.05). The average times to produce contours for AI, a urologist, and a radiologist were 96.5 seconds, 285.8 seconds, and 217.9 seconds, respectively (p<0.05 for each time).

Conclusion

The results suggest that, for the seminal vesicle, no statistical difference between the Dice similarities of AI-U, AI-R, and U-R was observed. Though the contours of the prostate and urethra did show a higher conformity for AI-U than AI-R and U-R, because AI-R and U-R presented similar conformities, it is unlikely this variability is linked to the AI’s accuracy. Hence, AI varies in conformity as much as other physicians and generates contours in less time, and current AI software holds potential as an effort-reducing tool for streamlining prostate cancer diagnostics in the clinical setting.

Funding

None

Lead Authors

Puranjay Shori, BS
UTHealth Houston School of Public Health

Jong Kim, MD
Doctors Imaging Group

Robert Carey, MD, PhD, FACS
Florida State University College of Medicine

Victoria Bird, MD
Urologic Integrated Care