Accurate diagnostic tissue segmentation and concurrent disease subtyping with small datasets.

Area of Science:

• Digital pathology
• Medical image analysis
• Computational oncology

Background:

• Accurate visual distinction and subtyping of diseased tissue in pathology slides are crucial for diagnosis and treatment.
• Current methods often require large datasets and complex models, limiting accessibility and efficiency.

Purpose of the Study:

• To develop a flexible, end-to-end platform for visual distinction and subtyping of diseased tissue in medical images, particularly pathology slides.
• To achieve high accuracy using small training datasets and easily shareable reduced-scale images.

Main Methods:

• An ensemble of lightweight convolutional neural networks (CNNs) was trained on subsets of reduced-scale whole-slide histopathology images (WSIs).
• A sequential workflow involving tile generation, classification by CNN ensembles, and mask creation was used for segmentation and subtyping.
• CNN predictions were combined to generate segmentation masks, which were then used to derive tiles for subtype classification.

Main Results:

• The approach was successfully applied to colorectal cancer (PAIP2020) and breast cancer metastasis (CAMELYON16) datasets.
• Segmentations achieved superior performance compared to more complex state-of-the-art models when evaluated using standard similarity metrics.

Conclusions:

• The developed platform offers an efficient and accurate method for analyzing large whole-slide histopathology images.
• This approach demonstrates the potential for improved diagnostic capabilities in digital pathology, even with limited data.