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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Surgical Oncology

Background:

• Radical prostatectomy difficulty assessment is crucial for patient outcomes.
• Current methods lack precision in predicting surgical complexity.
• Preoperative imaging offers potential for objective difficulty metrics.

Purpose of the Study:

• To develop and validate novel metrics for assessing radical prostatectomy difficulty using a two-stage deep learning approach.
• To leverage preoperative magnetic resonance imaging (MRI) for surgical planning and risk stratification.

Main Methods:

• A two-stage deep learning model (nnUNet_v2 and PointNet) was employed for prostate/pelvis segmentation and anatomical landmark localization from MRI.
• Novel metrics characterizing the spatial relationship between the prostate and pelvis were introduced.
• The model and metrics were validated on 290 patients from two cohorts (laparoscopic and robot-assisted radical prostatectomy).

Main Results:

• The deep learning pipeline achieved accurate segmentation (Dice 0.8641) and millimeter-level landmark localization.
• Specific spatial metrics (e.g., PSD2, PSD2×ρ) significantly correlated with Estimated Blood Loss and Operation Time.
• Validation on an external dataset confirmed the consistency and reliability of the findings.

Conclusions:

• The proposed two-stage deep learning method for anatomical landmark localization is feasible for surgical assessment.
• Spatial constraints between the prostate and pelvis are key indicators of radical prostatectomy difficulty.
• These novel metrics offer a promising avenue for improving preoperative surgical difficulty evaluation and patient management.