MegaSeg: Towards scalable semantic segmentation for megapixel images

Area of Science:

• Medical image analysis
• Computational pathology
• Deep learning for histopathology

Background:

• Megapixel image segmentation is vital for high-resolution histopathology analysis.
• Current GPU memory constraints necessitate patching and downsampling, compromising contextual information.
• Efficient segmentation of large-format images remains a significant challenge.

Purpose of the Study:

• Introduce MegaSeg, an end-to-end framework for semantic segmentation of megapixel histopathology images.
• Enable efficient processing of large images (e.g., 8192×8192 pixels) without sacrificing detail or context.
• Reduce memory usage in high-resolution image analysis.

Main Methods:

• Developed MegaSeg, an end-to-end framework utilizing streaming convolutional networks in a U-shaped architecture.
• Implemented a divide-and-conquer strategy for processing large images.
• Proposed the Attentive Dense Refinement Module (ADRM) within the decoder path to enhance local details and contextual information.

Main Results:

• MegaSeg enables efficient semantic segmentation of 67 MP images, preserving both global structure and local details.
• Demonstrated superior performance on public histopathology datasets.
• Achieved a significant improvement in the Free Response Operating Characteristic (FROC) score from 0.78 to 0.89 on the CAMELYON16 dataset when scaling input size from 4 MP to 67 MP.

Conclusions:

• MegaSeg effectively overcomes GPU memory limitations for megapixel image segmentation.
• The framework preserves essential global and local contextual information in high-resolution histopathology images.
• MegaSeg offers a promising solution for large-scale medical image analysis, enhancing diagnostic capabilities.