PDBL

PDBL enhances histopathological tissue classification to support semantic segmentation of whole-slide images and reduce reliance on pixel-level dense annotations.

Key Features:

• Backbone integration without re-training: PDBL augments the outputs of a well-trained classification backbone without requiring re-training.
• Multi-resolution image pyramid: For each image patch, PDBL constructs a multi-resolution image pyramid to capture pyramidal contextual information across scales.
• Deep-Broad block (DB-block): At each pyramid level, PDBL applies a DB-block to extract multi-scale deep and broad features.
• Classifier evaluations: PDBL has been integrated and evaluated with ShuffleNetV2, EfficientNetb0, and ResNet50.
• Reduced annotation requirement: PDBL reduces reliance on pixel-level dense annotations by improving tissue-level classification performance.

Scientific Applications:

• Limited-data classification: PDBL improves tissue-level classification performance when training data is limited.
• Dataset evaluations: Experimental results on the Kather Multiclass Dataset and the LC25000 Dataset show consistent performance gains across CNN backbones, with pronounced improvements for lightweight models trained on less than 10% of samples.

Methodology:

For each image patch PDBL constructs a multi-resolution image pyramid; at each pyramid level it applies a Deep-Broad block (DB-block) to extract multi-scale deep-broad features, and the extracted features are combined with an existing well-trained classification backbone without re-training.