Combining weakly and strongly supervised learning improves strong supervision in Gleason pattern classification.

Area of Science:

• Digital Pathology
• Computational Pathology
• Machine Learning in Medicine

Background:

• Training deep convolutional neural networks (CNNs) with whole slide images (WSIs) requires extensive, costly manual annotations.
• Weakly-supervised learning and transfer learning are strategies to mitigate data scarcity.
• Optimal methods for combining transfer learning with heterogeneous data sources and mixed annotation types (weak and strong) remain unclear.

Purpose of the Study:

• To evaluate CNN training strategies using transfer learning to combine weak and strong annotations from heterogeneous data sources.
• To explore the trade-off between classification performance and annotation effort in digital pathology.
• To assess fine-tuning strategies for CNNs trained on different data granularities.

Main Methods:

• Evaluating CNN training strategies based on transfer learning.
• Utilizing both strong (region) and weak (image-level) annotations.
• Fine-tuning a CNN model pre-trained on strong labels with weak labels from a different dataset.
• Comparing performance across tissue microarrays (TMAs) and WSIs.

Main Results:

• Model performance on strongly annotated data increased with more annotations, comparable to pathologists.
• Performance decreased significantly when applied directly to WSI classification.
• Fine-tuning with weak WSI labels improved Gleason scoring task performance.
• Combining weak and strong supervision enhanced classification of Gleason patterns.

Conclusions:

• Combining weak and strong supervision effectively improves classification performance in digital pathology.
• Transfer learning, when fine-tuned on the target dataset (WSI), yields optimal downstream models.
• Strategies for training CNNs with limited annotated data and heterogeneous sources were developed.
• Source code is available for reproducibility.