Ventricle Detection Case Study: Comparison of Current Deep Learning Architecture Performances

The brain is one of the body's most critical structures, governing or controlling all organs. Therefore, damage to the brain can have life-threatening effects. Midline shift (MLS), on the other hand, is a serious condition that results from the migration of brain compartments from one to the other for various reasons, leading to various neurological symptoms such as increased cranial pressure, impaired consciousness, paralysis, and even brain death. Therefore, early diagnosis is crucial in cases of MLS. The first step in diagnosing MLS is identifying the ventricles, one of the structures that define the midline in the brain. The ventricles can be visualized in different ways depending on the case, increasing the risk of physicians making errors in the initial diagnosis. The aim of this study is to create an artificial intelligence (AI)-powered ventricle detection architecture that will assist physicians in minimizing this error. For this purpose, Faster Region Based Convolutional Networks (Faster R-CNN), You Only Look Once X (YOLOX), Cascade Region Based Convolutional Neural Network (Cascade R-CNN), Efficient Detection (EfficientDet), enhanced YOLO versions (YOLOv8n, YOLO11n) and current transformer-based deep learning (DL) architecture Shifted Window Transformer (Swin Transformer) were tested. The study was conducted on the mainly CQ500 dataset and results were validated with a subset of tests obtained from the RSNA brain hemorrhage dataset. Both datasets are open and marked with a bounding box technique. Comparisons of current DL models reveal which models are more advantageous in which situations. Furthermore, the inclusion of an open dataset provides a basis for future research.