Localization of mixed intracranial hemorrhages by using a ghost convolution-based YOLO network

• Mixed ICH is a serious health disease, so an efficient light-weight multi-scale YOLO-GCB architecture is given for the localization of each hemorrhage in the given CT. • A novel mosaic training method is used to boost the performance by creating a greater number of mixed hemorrhage cases than in the original dataset. • Additional developments of ghost convolution and C3 ghost modules improves the speed by reducing the number of computations. • Memory required to deploy the proposed model either in cloud or in embedded devices is less by comparing with the other state-of-the-art existing models by producing similar results with respect to other metrics. • Finally, the mixed hemorrhages problem is eliminated by predicting a bounding box around each hemorrhage along with a class name and confidence score. Intracranial hemorrhage (ICH) is a serious medical condition that must be diagnosed in a stipulated time through computed tomography (CT) imaging modality. However, the neurologist must initially confirm the specific type of hemorrhage to prescribe an effective treatment. Although conventional image processing and convolution-based deep learning models can effectively perform multiclass classification tasks, they fail to classify if a CT input image contains multiple hemorrhages in a single slice and takes a lot of time to make the final predictions. To overcome these two difficulties, we proposed a novel YOLOv5x-GCB model that can be able to detect multiple hemorrhages with limited resources by employing a ghost convolution process. The advantage of ghost convolution is that it produces the same number of feature maps as vanilla convolution while using less expensive linear operations. Another feature of the proposed model is that it uses the mosaic augmentation technique throughout the training to improve the accuracy of mixed hemorrhage detection. A brain hemorrhage extended dataset containing 21,132 slices from 205 positive patients was used in training and validating the proposed model. To test the robustness of the proposed model, we created a separate dataset with the existing segmentation data, which are available in PhysioNet. As a result, the proposed model achieved an overall precision, recall, F1- score, and mean average precision of 92.1%, 88.9%, 90%, and 93.1%, respectively. In addition to these metrics, other parameters were used in evaluating the proposed model and checking its lightweight capability in terms of memory size and computational time. Results showed that our proposed model can be used in real-time clinical diagnosis by using either embedded devices or cloud services.