Classification and Localization of Intracranial Hemorrhages Through Deep Learning Based Optimization Algorithms

Abstract

Intracranial hemorrhage (ICH) is a critical medical condition requiring prompt diagnosis and intervention. Traditional methods of detecting and localizing ICH, such as manual examination of CT scans, are time-consuming and subject to inter-observer variability. Identifying and localizing various subtypes of ICH through deep learning (DL) techniques is increasingly important in the medical domain. The present thesis effectively tackles the intricacy and fluctuations present in various hemorrhagic lesion sizes and kinds through the utilization of advanced DL models, namely CNN, and their variations. The first work required an easy-to-implement algorithm that could distinguish between normal and ICH brain CT scans; hence a basic CNN technique was adopted. The main drawbacks of this method are its short dataset and its disregard for inter-pixel dependence even when accounting for CT input slices. Bidirectional Long short term Memory (Bi-LSTM) and Bidirectional Gated recurrent unit (Bi-GRU) modules are used to create a stacked design that addresses these two issues. A transfer learning-based DenseNet121 architecture was utilized to carry out the multi-class categorization. newlineIn the second work, a framework called YOLOv5x-GCB was designed to localize the mixed ICH with minimal resources. YOLOv5, known for its speed and accuracy in object detection tasks, is adapted to handle the complexity of medical imaging data. The lightweight nature is achieved by introducing the concept of ghost convolution. This idea significantly decreased the number of computations needed. The primary advantage of this module in the suggested model is that it stores the weights in about one-third less memory than the existing model does. The main drawback is that a slight reduction in detection accuracy occurred because of the minor ICH classes not being detected. To overcome this drawback, the third work presents an attention-based transformer strategy to introduce the GEL-TTA Net.