Patch relevance estimation and multilabel augmentation for weakly supervised histopathology image classification.

Area of Science:

• Digital Pathology
• Machine Learning
• Computational Biology

Background:

• Weakly supervised learning (WSL) is crucial for histopathological image analysis, using image-level diagnoses as weak labels for fixed-size patches.
• Multiclass classification in histopathology is challenging due to co-existing diagnostic categories within a single image, leading to label uncertainty.

Purpose of the Study:

• To address label uncertainty in multiclass histopathological image analysis using weakly supervised learning.
• To develop an improved patch-based WSL method for accurate classification of complex histopathology images.

Main Methods:

• A two-branch architecture estimating patch-level class likelihoods and relevance weights.
• A complementary training strategy combining outputs for image-level predictions.
• A multilabel augmentation strategy creating new training samples from pairs of images to enrich the dataset.

Main Results:

• The proposed method outperforms conventional weakly supervised approaches on multiclass breast histopathology datasets.
• Demonstrated improvements in classification accuracy and robustness, especially for underrepresented diagnostic classes.
• Effective modeling of the relationship between image-level labels and patch-level content.

Conclusions:

• The novel architecture and multilabel augmentation effectively improve learning under label uncertainty in histopathology.
• The approach enhances diagnostic accuracy and robustness in complex multiclass scenarios.
• Potential for developing more accurate and scalable diagnostic support systems in digital pathology.