Generating 2.5D pathology for enhanced viewing and AI diagnosis

Area of Science:

• Digital Pathology
• Computational Anatomy
• Medical Imaging Analysis

Background:

• Histological analysis of biopsy samples involves evaluating complex 3D tissue structures across multiple slides, a process that is time-consuming due to manual zooming and panning.
• Current deep learning models often analyze 2D cross-sections, failing to capture crucial 3D spatial information inherent in biopsy specimens.

Purpose of the Study:

• To develop a novel framework for constructing 2.5D biopsy cores from serial tissue sections.
• To enable enhanced visualization for pathologists and provide suitable input for advanced deep learning models, such as video transformers, for capturing depth-wide spatial dependencies.

Main Methods:

• A novel morphology-preserving alignment framework was developed to extract and co-align serial tissue sections, creating 2.5D biopsy cores.
• The framework was applied to construct 2.5D cores for a large cohort of prostate, breast, and renal biopsies.
• Downstream task evaluation included training a deep learning-based cancer grading model and conducting a pathologist reader study using the 2.5D cores.

Main Results:

• The framework successfully constructed 2.5D biopsy cores from over 10,000 prostate biopsies, along with breast and renal biopsies.
• The 2.5D cores demonstrated utility in downstream tasks, including the development of a deep learning cancer grading model.
• Pathologist reader studies indicated the potential for enhanced viewing and analysis using the 2.5D core representation.

Conclusions:

• The developed framework effectively generates 2.5D biopsy cores, preserving tissue morphology and enabling the capture of 3D spatial information.
• These 2.5D cores offer a valuable tool for both pathologist visualization and advanced computational analysis in digital pathology.
• This approach represents a significant advancement in leveraging 3D tissue information for improved diagnostic accuracy and machine learning applications in histopathology.