Intracranial Hemorrhage Grading using Novel ConceptionNet

Intracranial hemorrhage is a fatal disease with severe outcomes. Non-contrast computed tomography (CT) scan is the preferred method of evaluating patients for hemorrhage. Traditional visual evaluation is time-consuming and reader dependent. This study is aimed at developing an automated algorithm for hemorrhage detection and its sub-type classification using deep learning. However, unavailability of a large tagged dataset for training leads to complicated problems. Recently, pre trained networks offered superior performance in classification tasks and after fine-tuning these algorithms outperformed traditional algorithms. We hypothesize that on CT images acute hemorrhage has hyper-dense signal intensity and deep learning algorithms are capable of detecting and grading hemorrhage. In this work, state-of-the-art transfer learning approaches analyse images with frozen pre-trained weights. The final layers are trained on hemorrhage images and the most probable class is predicted. Moreover, we propose a novel algorithm for the hemorrhage detection and sub-type classification. The algorithm was trained on 2848 CT images of 46 male and 36 female subjects. The proposed methodology included cutting-edge pre trained models: the Inception and EfficientNetB0 as well as the novel ConceptionNet. The area under the curve for epidural, intraparenchymal, intraventricular, subarachnoid and subdural are 0.983, 0.973, 1.0, 0.968 and 1.0 respectively. The classification accuracy of EfficientNetB0, InceptionV3 and ConceptionNet models are 0.96, 0.95 and 0.96. The algorithm accomplished excellent performance giving confidence for its application as a tool to assist junior radiologists.