Harnessing clinical annotations to improve deep learning performance in prostate segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Developing high-quality annotated datasets for medical AI is costly and time-consuming.
• Clinically generated annotations often lack the precision required for research-grade models.
• This study explores leveraging large, less-refined datasets to improve segmentation model performance.

Purpose of the Study:

• To evaluate the effectiveness of using a large dataset with clinical annotations for training segmentation models.
• To assess the performance enhancement on smaller, research-quality datasets when using a pre-trained model.
• To determine the minimum data proportion required for effective template training.

Main Methods:

• Trained a 3D U-Net convolutional neural network (CNN) on a large dataset (1,620 segmentations).
• Used the trained model as a template to fine-tune on two challenge datasets (PROMISE12, ProstateX-2).
• Evaluated performance using Dice scores and Hausdorff distance with five-fold cross-validation, testing various data proportions and an out-of-domain dataset.

Main Results:

• The template model achieved state-of-the-art performance on the large dataset (Dice 0.916).
• Fine-tuning with the template significantly improved performance on challenge datasets (30% and 49% Dice enhancement).
• Performance gains were observed even with only 5% of the original training data.

Conclusions:

• A state-of-the-art segmentation model can be trained effectively using unrefined clinical prostate annotations.
• This template model significantly enhances performance on external prostate segmentation tasks.
• Leveraging large, clinically annotated datasets is a viable strategy for developing high-performance medical AI models efficiently.