PDBL: Improving Histopathological Tissue Classification With Plug-and-Play Pyramidal Deep-Broad Learning.

Area of Science:

• Digital pathology
• Computer-aided diagnosis
• Computational imaging

Background:

• Histopathological tissue classification simplifies semantic segmentation of whole slide images, reducing the need for pixel-level annotations.
• Current methods primarily use Convolutional Neural Network (CNN) classification backbones for this task.

Purpose of the Study:

• To introduce Pyramidal Deep-Broad Learning (PDBL), a lightweight, plug-and-play module designed to boost classification performance of existing CNN backbones.
• To improve histopathological tissue classification without requiring extensive re-training.

Main Methods:

• PDBL constructs multi-resolution image pyramids for each patch to capture pyramidal contextual information.
• It extracts multi-scale deep-broad features using a novel Deep-Broad block (DB-block).
• The module was integrated with ShuffLeNetV2, EfficientNetb0, and ResNet50 backbones and evaluated on the Kather Multiclass and LC25000 datasets.

Main Results:

• PDBL consistently enhanced tissue-level classification performance across different CNN backbones.
• The module showed particular effectiveness with lightweight models and limited training samples (under 10%).
• Significant savings in computational resources and annotation efforts were observed.

Conclusions:

• PDBL is an effective and efficient module for improving histopathological tissue classification.
• It offers a valuable solution for scenarios with limited annotated data and computational constraints.
• The plug-and-play nature allows easy integration with various pre-trained classification models.