Multi-resolution consistency semi-supervised active learning framework for histopathology image classification

Histopathology image classification is one of the most important fundamental tasks in the automation analysis of whole slide imaging and is essential for computer-aided pathological diagnosis. Training a deep learning model for analyzing histopathology images highly depends on a large amount of high-quality labeled data. However, there are always conflicts between the quality and quantity of labels. To reduce the dependency on labels, we propose a multi-resolution consistency semi-supervised active learning (MCSSAL) framework for histopathology image classification. This framework first introduces two distinct consistency regularizations for semi-supervised learning (SSL) method. One is the consistency between a weakly perturbed image and strongly perturbed images with multi-resolutions. The other one is the consistency among strongly perturbed images at various resolutions. It can be helpful to fully extract the underlying pathological features and complementary information from images with different resolutions during training. Moreover, we design an active learning strategy to select hard-to-learn samples for manual labeling to enhance the performance of SSL. Specifically, this strategy effectively evaluates the learning difficulty of samples by combining multi-resolution inconsistency and uncertainty measures that are coherent with the training objective of SSL. This framework integrates the advantages of semi-supervised learning and active learning to make them more effective with a mutual collaboration method. The MCSSAL framework was performed on three histopathology image datasets. Results showed that MCSSAL obtained superior results in classification compared to other state-of-the-art methods. Therefore, our MCSSAL framework can potentially make pathological image classification more efficient only using limited labels. © 2024 Elsevier Ltd