A Multiple-Instance Learning-Based Convolutional Neural Network Model to Detect the IDH1 Mutation in the Histopathology Images of Glioma Tissues

The IDH1 mutation is the most frequent somatic mutation in gliomas, and it has an important impact on the treatment outcome of gliomas. Clinically, the gold standard methods for the IDH mutation detection are the immunohistochemistry and gene sequencing techniques, whereas using the histopathology images of the glioma tissues for IDH mutation identification has not been reported. In this study, we propose a convolutional neural network (CNN) model that is trained on histopathology images of glioma samples using multiple instance learning (MIL), which links the benefits of the end-to-end classification power of the deep neural network with the MIL by aggregating the scores of the instances to the bag-level score. The attention layer is also implemented to facilitate the performance of the MIL aggregation. The results show that our MIL-based CNN model has achieved good performance in the classification of the IDH1 mutation in the glioma images, with the area under the curve of 0.84. Besides, several image segmentation strategies, CNN architectures, and MIL pooling operators have been implemented and analyzed to investigate the effect of these settings on the model performance. To our knowledge, it is the first study to identify the IDH1 mutation by using the histopathology images of the glioma tissues, providing a novel and insightful method for glioma IDH mutation diagnosis.