Shallow 3D CNN for Detecting Acute Brain Hemorrhage From Medical Imaging Sensors

Successive layers in convolutional neural networks (CNN) extract different features from input images. Applications of CNNs to detect abnormalities in the 2D images or 3D volumes of body organs have recently become popular. However, computer-aided detection of diseases using deep CNN is challenging due to the absence of a large set of training medical images/scans and the relatively small and hard to detect abnormalities. In this paper, we propose a method for normalizing 3D volumetric scans using the intensity profile of the training samples. This aids the CNN by creating a higher contrast around the abnormal region of interest in the scan. We use the CQ500 head CT dataset to demonstrate the validity of our method for detecting different acute brain hemorrhages such as subarachnoid hemorrhage (SAH), intraparenchymal hemorrhage (IPH), subdural hematoma (SDH), and intraventricular hemorrhage (IVH). We compare the proposed method with a baseline, two variants of the 3D VGGNet architectures, Resnet, and show that the proposed method achieves significant improvement in classification performance. For binary classification, we achieved the best F1 score of 0.96 (normal vs SAH), 0.93 (normal vs IPH), 0.98 (normal vs SDH), and 0.99 (normal vs IVH), and for four-class classification, we obtained an average F1 score of 0.77. Finally, we show a limitation of the proposed method while detecting varied abnormalities. The proposed method has applications for abnormality detection for different organs.