Breast cancer has become a worldwide disease in recent years. However, despite its growing prominence, the number of pathologists equipped to handle these cases is insufficient. Computer-aided diagnosis (CAD) system contributes to reduce costs and improve efficiency of this process. A framework based on convolutional neural networks (CNNs) which could be used to automatically detect the multi-class cancer areas on gigapixel pathology slide images was proposed. Moreover, combining the slide image characters, rescale and careful data augmentation methods were used to train the patch-based model with a small dataset. To validate the developed framework, we conducted experiments with Breast Cancer Histology Challenge (BACH) dataset and obtained International Conference on Image Analysis and Recognition (ICIAR) score of 0.582, outperforming the second-place finisher in BACH2018, for the 4-class tissue segmentation task.
Malignant melanoma (MM) of the eyelid is of high malignancy, high mortality, and easy to metastasize. Currently, the gold standard for MM treatment and prognosis is histopathology, but the diagnosis of different experts is often divergent. The computer-aided diagnosis based on deep learning helps to improve efficiency and accuracy. In this paper, a complete set of methods for MM diagnosis is proposed using the convolutional neural network (CNN) to classify the patch level pathological images. Hematoxylin and Eosin (H and E)-stained pathological images of the eyelids are classified as malignant melanoma and non-malignant melanoma (NMM). The prediction results are filled by location in the probabilistic map of the whole slide image level. Random forest classifier based on CNN inference results extract 31- dimensional features to achieve whole slide image-level classification. The color constancy method and the edge extraction mapping method based on the Sobel operator (EMBS) can significantly improve the performance of the model. The patch level classification results show that the balance accuracy is 93% on the Second Affiliated Hospital, Zhejiang University School of Medicine (ZJU-2) test set, and the balance accuracy is 89.4% on the Shanghai Ninth People’s Hospital, Shanghai JiaoTong University School of Medicine (SJTU) test set. The corresponding area under curve (AUC) is 0.990 and 0.970. For whole slide image level classification results, the AUC for SJTU test set is 0.999, the sensitivity is 100%, and the specificity is 97.4%. As a result, our model can effectively tackle the challenge of clinicopathological diagnosis and relieve the pressure of pathologists.