Comparison of Intracranial Hemorrhages Detection Performances of Deep Learning Models on CT Images

Early detection of Intracranial hemorrhage (ICH) is crucial to prevent life-threatening conditions and mortality. Computed Tomography (CT) is the primary imaging system used to diagnose intracranial hemorrhages, which provides distinct differences between epidural, subdural, and subarachnoid pleurae. However, the occurrence of ICH could be indistinct and might cause misdiagnosis. Recent advancements in artificial Intelligence (AI) and deep learning technologies provide effective and rapid analysis of low and high-level features of images, even though the appearance of the region of interest is indistinct. This efficacy of the deep learning methods is also improved by transferring the knowledge obtained in another domain to different tasks. In this paper, we implemented three state-of-the-art pretrained deep learning models, EfficientNetB0, DenseNet201, and ResNet101, using a transfer learning approach to detect ICH in order to help and assist medical doctors. The experiments are performed using 5-fold cross-validation, and the predictive results are analyzed using seven evaluation metrics. The DenseNet201 outperformed ResNet101 and EfficientNet B0 models in all evaluation metrics and achieved 0.8076, 0.8451, and 0.981 Sensitivity, F1, and ROC AUC scores. The results showed that deep learning models can be used to detect ICH accurately; however, further improvements are required to increase the sensitivity.