Intracranial Hemorrhage Detection And Category Classification Using Attention Aware Pure Convnets

In this paper, we propose a methodology for detection and category classification of intracranial hemorrhages. The contributions of this study are two-fold, 1) we propose a Channelwise Attention induced pure convolutional neural network architecture (CA-ConvNeXt), 2) we propose a novel angularmargined focal loss to train our network on the highly imbalanced dataset where the incident rate is about 1%. This loss helps in emphasizing the under-represented categories and increases the margin between the features in the hypersphere. We train our network in two stages. The first is the screening stage which detects the presence of hemorrhage. The abnormal classified image from this stage is sent to the second stage which is used for hemorrhage category classification. We utilize the extensive publicly available RSNA ICH detection challenge and PhysioNet datasets to illustrate the performance of our proposed method. This methodology is tested on a 10% hold-out dataset resulting in sensitivity of 99% and precision of 91% on the screening stage(stage-1), an average sensitivity of 93% and average precision of 92% in hemorrhage category classification(stage-2). The best AUROC of 0.97 was achieved on the Subarachnoid hemorrhage across both datasets. The challenging categories are Epidural hemorrhage because the data is only 3% of the total abnormal images and the Subdural hemorrhage because the hemorrhage cannot be visualized within one single window.