Segment-then-Segment: Context-Preserving Crop-Based Segmentation for Large Biomedical Images

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Large medical image file sizes pose memory challenges for training machine learning models.
• Downsampling medical images to reduce size leads to significant information loss, impacting model accuracy.
• Semantic segmentation in medical imaging requires high-resolution details for accurate analysis.

Purpose of the Study:

• To present a generalizable approach, Segment-then-Segment, for training semantic segmentation neural networks with reduced input sizes.
• To overcome the limitations of downsampling by preserving full-resolution details through image cropping.
• To improve segmentation performance, particularly pixel-wise recall, in medical imaging tasks.

Main Methods:

• The Segment-then-Segment approach utilizes image crops instead of downscaling to manage input size.
• An initial segmentation network processes a downscaled image.
• Salient crops from the full-resolution image, with context, are segmented by a second specialized network, and masks are merged.

Main Results:

• The Segment-then-Segment method significantly enhances segmentation performance on small network input sizes.
• The approach demonstrates superior results compared to baseline models trained on downscaled images.
• Improvements are particularly notable in pixel-wise recall across various medical image modalities.

Conclusions:

• Segment-then-Segment offers an effective strategy for training semantic segmentation models on large medical images without compromising performance.
• The cropping-based method preserves crucial image details lost in downscaling, leading to better segmentation accuracy.
• This approach is broadly applicable to diverse medical imaging modalities, including microscopy, colonoscopy, and CT scans.