Semisupervised Training of a Brain MRI Tumor Detection Model Using Mined Annotations

Area of Science:

• Artificial intelligence in medical imaging
• Machine learning for oncology
• Radiomics and computational pathology

Background:

• Supervised training for AI cancer imaging requires extensive manual annotation, which is time-consuming and expensive.
• Existing clinical data within Picture Archiving and Communication Systems (PACS) represents a largely untapped resource for AI model development.
• Automated tumor detection is a foundational AI task in medical imaging.

Purpose of the Study:

• To assess the feasibility of using data mined from PACS for semisupervised training of brain MRI tumor detection models.
• To investigate the effectiveness of automatically curating and utilizing existing clinical image annotations.
• To improve the performance of AI models for detecting brain tumors in MRI scans.

Main Methods:

• Retrospective analysis of brain MRI scans (2012-2017) with existing annotations from PACS.
• Conversion of line annotations to bounding boxes and application of inclusion criteria.
• Supervised training of object detection models (RetinaNet, Mask R-CNN) using mined data.
• Implementation of a self-labeling strategy to expand the training dataset.
• Evaluation of model performance using a held-out test set of 754 manually labeled images.

Main Results:

• Over 11,880 bounding boxes were extracted from PACS annotations for initial model training.
• Initial models achieved high F1 scores (RetinaNet: 0.886, Mask R-CNN: 0.908).
• Semisupervised learning via self-labeling significantly improved F1 scores to 0.935 and 0.954 (p < .001).

Conclusions:

• Semisupervised learning applied to mined PACS annotations substantially enhances AI-driven tumor detection accuracy.
• This approach repurposes existing data silos, offering a cost-effective pipeline for AI development across imaging modalities.
• The developed method shows potential for automated tumor detection in various radiological applications.