Label Diffusion Graph Learning network for semi-supervised breast histological image recognition

Abstract

Breast cancer has been a significant cause of female death worldwide, usually recognized by histological images. Due to the outstanding performance of computer-aided methods, many breast histological recognition algorithms are based on deep learning. Nonetheless, these advanced breast cancer recognition approaches exist two primary issues. Firstly, these methods rely on abundant pre-annotated labels, which is an expensive and heavy workload. Besides, they lack the multi-level (both pixel-level and sample-level) feature learning ability to reveal the complex semantic characteristics in histology tissue images. For these challenging problems, we proposed a Label Diffusion Graph Learning (LDGL) method, which can optimize the model in a semi-supervised manner with limited labels, and then adopt graph convolution layers to mine correlations among different breast tissues. Concretely, our LDGL model first employs a convolutional neural network and graph convolution network to extract comprehensive multi-level representations. Then it improves the identification capability through a multi-level consistency loss among unlabeled breast images. Finally, a novel pseudo-label building method is introduced to combine confidence, graph degree, and the expected calibration error to expand labeled data via credible breast samples. To prove the excellent recognition performance of our LDGL model, we compare it with several superior breast cancer histological image recognition approaches. The results indicate that our model (with only 20% labels) achieves 95.05% Accuracy, 95.27% Precision, 95.05% Recall, and 95.0% F1-score, separately. Moreover, we also verify the contribution of each module in LDGL through a series of ablation experiments. In conclusion, the evaluation proves that our LDGL model is excellent and robust for semi-supervised breast histological image recognition task.