Deep Learning for Predicting Difficulty in Radical Prostatectomy: A Novel Evaluation Scheme

Abstract

Objective

To explore new metrics for assessing radical prostatectomy difficulty through a two-stage deep learning method from preoperative magnetic resonance imaging.

Methods

The procedure and metrics were validated through 290 patients consisting of laparoscopic and robot-assisted radical prostatectomy procedures from two real cohorts. The nnUNet_v2 adaptive model was trained to perform accurate segmentation of the prostate and pelvis. A modified network PointNet was used for indirectly regressing 15 anatomical landmarks based on Gaussian heatmaps. Novel metrics proposed in this study that characterized the spatial relationship between the prostate and pelvis were included to evaluate the surgical difficulty.

Results

The two-stage process achieved decent segmentation and landmark localization results with the mean validation dice of 0.8641 and millimeter-level accuracy. We found the coefficients of PV, ρ, PT, PAP, AG, PSD₁, PSD₂, πρ²/ISTA, AG+PG, AG$\times$PG, PSD₂$\times$ρ, and PAP/(AG+PG) with Estimated Blood Loss and PSD₂, PSD₂$\times$ρ with Operation Time, respectively, with statistic significant, which provides possibilities for assessing surgical difficulty evaluation. The entire pipeline had been validated on the external dataset, and the results were consistent.

Conclusion

The two-stage anatomical landmark localization approach is feasible. Indicators describing pelvic-prostate spatial constraints significantly impact surgical difficulty in radical prostatectomy, leading to increased blood loss and longer operation times, while isolated pelvic measurements have minimal effect on surgical outcomes.