Fast, light, and scalable: harnessing data-mined line annotations for automated tumor segmentation on brain MRI

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Neuro-oncology imaging

Background:

• Fully supervised learning for brain MRI tumor segmentation requires extensive manual annotation, limiting practical application.
• Developing automated segmentation models is crucial for efficient and scalable clinical workflows.

Purpose of the Study:

• To investigate the utility of mined line annotations for brain MRI tumor segmentation model development.
• To develop an automated pipeline for generating accurate tumor segmentation without manually segmented training data.

Main Methods:

• Leveraged a tumor detection model trained on clinical line annotations from PACS to generate pseudo-masks for enhancing tumors on T1-weighted post-contrast MRI.
• Employed a semi-supervised learning (SSL) framework with iterative self-refinement to improve pseudo-mask quality.
• Compared conventional full-image segmentation with a hybrid method combining full-image and patch segmentation.

Main Results:

• Baseline segmentation models (U-Net, Mask R-CNN, HRNet) improved with self-refinement, achieving Dice scores up to 0.873.
• The hybrid inference method demonstrated superior performance with a Dice score of 0.884 (Mask R-CNN).

Conclusions:

• Clinical line annotations mined from PACS can be effectively utilized to develop accurate brain MRI tumor segmentation models.
• This automated pipeline offers a rapid and efficient method for establishing tumor segmentation capabilities across various radiology modalities.
• The findings eliminate the need for manually segmented training data, overcoming a significant bottleneck in model development.