Continuous Refinement-based Digital Pathology Image Assistance Scheme in Medical Decision-Making Systems.

Area of Science:

• Digital pathology
• Computational pathology
• Medical image analysis

Background:

• Digital pathology images offer rich cellular data for tumor diagnosis, aided by computer-aided diagnostics.
• Current cascade-based models struggle with ultra-high resolution images, leading to computational costs and information loss.
• Existing methods require downsampling and cropping, compromising cellular details and global context.

Purpose of the Study:

• To develop an improved computer-aided diagnostic system for digital pathology.
• To address the limitations of cascade-based models in handling high-resolution pathology images.
• To enhance the accuracy and efficiency of tumor diagnosis using digital pathology images.

Main Methods:

• Proposed a Digital Pathology Image Assistance Program (CRSDPI) based on continuous improvement.
• Utilized the maximum inter-class variance method for region of interest localization.
• Developed a two-phase continuously refined segmentation network (TCRNet) combining a coarse segmentation network and an enhanced continuous refinement model.
• Incorporated an auxiliary loss term for faster convergence and an implicit function to reduce computational cost and reconstruct details.

Main Results:

• The TCRNet model refines targets by aligning features without cascading decoder operations.
• Demonstrated superior prediction accuracy compared to existing methods.
• Achieved significant improvements in computational speed for image analysis.
• Successfully applied to digital pathology images of breast cancer and osteosarcoma.

Conclusions:

• The proposed TCRNet model offers a more efficient and accurate approach to digital pathology image analysis.
• CRSDPI enhances medical decision-making systems by providing reliable tumor diagnostic support.
• The method effectively overcomes the limitations of traditional cascade-based models in high-resolution image processing.