Reinforcement learning-guided segment anything model for MRI prostate and dominant intraprostatic lesions...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiation Therapy Planning

Background:

• Accurate segmentation of prostate and dominant intraprostatic lesions (DILs) on MRI is crucial for prostate cancer radiation therapy.
• Challenges in DIL segmentation include small datasets, institutional bias, and variable imaging protocols.
• Existing methods often rely on manual prompts, limiting automation.

Purpose of the Study:

• To develop a fully automated pipeline for segmenting the prostate and DILs on MRI.
• To integrate a localization network with a fine-tuned Segment Anything Model (SAM).
• To evaluate the performance of the automated segmentation pipeline on multi-institutional datasets.

Main Methods:

• A two-stage pipeline was developed: (1) a reinforcement learning-based localization network for bounding box prediction, and (2) a fine-tuned SAM model for segmentation.
• Utilized two large datasets: PI-CAI (1,476 patients) and The Cancer Imaging Archive (803 patients).
• Performance was assessed using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), and detection rates, with analysis stratified by lesion volume.

Main Results:

• The pipeline achieved high accuracy for prostate segmentation (mean DSC: 0.896, mean IoU: 0.818).
• For DIL segmentation, the mean DSC was 0.592 and mean IoU was 0.446, with an 89% detection rate.
• Performance remained robust across different lesion volumes, indicating generalizability.

Conclusions:

• A fully automated framework for prostate and DIL segmentation on MRI was successfully developed.
• The integrated approach demonstrates robust performance across diverse datasets and lesion characteristics.
• This method holds significant potential for enhancing clinical workflows in prostate cancer radiation therapy planning.