Promptable segmentation with region exploration enables minimal-effort expert-level prostate cancer delineation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate prostate cancer segmentation on MRI is vital for image-guided interventions.
• Challenges include subtle tumor appearances, protocol variations, and limited expert availability.
• Current automated methods require extensive, often inconsistent, annotations, while manual segmentation is time-consuming.

Purpose of the Study:

• To develop a framework for accurate prostate cancer segmentation using minimal user annotation effort.
• To bridge the gap between manual and fully automated segmentation methods.
• To enable efficient and adaptive cancer delineation through a novel approach.

Main Methods:

• A framework combining reinforcement learning (RL) with user-prompted region growing.
• Iterative refinement of segmentation masks guided by an RL agent observing image data and current segmentations.
• A reward system balancing accuracy and uncertainty to optimize segmentation and escape local optima.

Main Results:

• The framework achieved performance comparable to manual radiologist segmentation on two public datasets (PROMIS, PICAI).
• Outperformed previous best automated methods by 9.9% (PROMIS) and 8.9% (PICAI).
• Reduced annotation time by tenfold compared to manual segmentation.

Conclusions:

• The proposed framework achieves radiologist-level prostate cancer segmentation with a fraction of the annotation effort.
• Highlights the potential of RL for adaptive and efficient cancer delineation.
• Provides an open-source implementation for further research and application.