In-Domain Transfer Learning Strategy for Tumor Detection on Brain MRI

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Transfer learning is crucial for data-scarce domains like medical imaging.
• The efficacy of natural image datasets for pre-training in medical fields remains debated.
• Quantitative comparison of transfer learning strategies is needed for medical object detection.

Purpose of the Study:

• To quantitatively compare transfer learning for medical object detection using natural vs. medical image datasets.
• To evaluate different weight initialization methods for transfer learning.
• To assess the impact of data augmentation on transfer learning performance.

Main Methods:

• Utilized Mask R-CNN architecture for object detection and segmentation.
• Employed MS COCO (natural) and BraTS 2020 (medical) datasets for pre-training.
• Gazi Brains 2020 dataset served as the target for medical object detection.
• Compared five different weight initialization strategies.

Main Results:

• Medical image pre-training yielded 10% higher success and 24% better convergence than natural image pre-training.
• Transfer learning with MS COCO or random weights showed performance similar to data augmentation.
• In-domain transfer learning demonstrated higher efficiency than cross-domain transfer learning, irrespective of data volume.

Conclusions:

• In-domain transfer learning, particularly with medical datasets, is superior for medical object detection tasks.
• The Gazi Brains 2020 dataset is introduced for in-domain transfer learning in brain MRI analysis.
• Deep neural networks benefit from knowledge transfer within the same domain for improved medical image analysis.