PathAL: An Active Learning Framework for Histopathology Image Analysis.

Area of Science:

• Digital Pathology
• Computational Pathology
• Medical Image Analysis

Background:

• Deep neural networks, especially convolutional networks, are widely used for histopathology image analysis.
• Training these models requires extensive expert-annotated datasets, which are costly and time-consuming to create.
• Variability in expert annotations can lead to imperfect labels, impacting model performance.

Purpose of the Study:

• To present PathAL, a novel active learning (AL) framework designed to minimize the number of expert annotations required for histopathology image analysis.
• To improve the efficiency and accuracy of training deep learning models in digital pathology.
• To address the challenges of noisy labels and expert variability in histopathology datasets.

Main Methods:

• PathAL employs a dual-data selection strategy in each iteration: identifying informative samples for expert annotation and confident pseudo-labeled samples for automatic inclusion in the training set.
• The framework systematically identifies and excludes noisy samples while retaining challenging (hard) samples to enhance model generalization.
• A heuristic approach is used to differentiate between noisy and hard samples, optimizing the training data composition.

Main Results:

• PathAL achieved comparable performance to fully supervised learning on a prostate cancer Gleason grading task, but with 40% fewer expert annotations.
• The framework demonstrated improved classification performance by strategically selecting informative and confident samples.
• Ablation studies confirmed the effectiveness of individual components within the PathAL framework.

Conclusions:

• The proposed PathAL framework significantly reduces the annotation burden in histopathology image analysis while maintaining high model performance.
• PathAL offers an effective solution for training deep learning models with limited labeled data, addressing the issue of noisy labels.
• The framework shows promise for advancing automated analysis in digital pathology, validated by both computational experiments and pathologist studies.