Aneurysm Detection from Brain CT Scans

One of the primary subjects of human research has always been health. The first application of artificial intelligence in the medical field dates back a long time, but particularly in recent years, with the aid of new, improved AI techniques and methods, several useful machine learning models have been developed that are currently being used in hospitals and other medical facilities. Brain Aneurysms are among the diseases that are particularly difficult to detect. They are extremely difficult to identify in the first place, and the diagnostic procedures are expensive, time-consuming, and difficult to complete. On the other hand, even late discovery of this issue can be extremely fatal or cause patients to sustain permanent disabilities. The primary goal of this study was to aid the medical community in finding Aneurysms in a much quicker and easier manner. It was intended to create and set up a classification model that could categorize each slice of a typical brain ct scan as either healthy or unwell (containing Aneurysm). The primary concerns with this assignment were that, in most cases, it is quite challenging to identify brain Aneurysms from other brain tissues in brain CT scans. This is one of the main problems with this task. The second issue was that we only had a dataset of 50 patients, and only 25 of them displayed symptoms of a brain Aneurysm so overall the dataset was small and highly unbalanced towards healthy slices. To solve these issues, we initially carried out several data cleaning and preparation methods to make the most of the dataset that we had. In order to supplement the limited dataset we had, we also used one of the RSNA public datasets, which had classification labels for many patients' hemorrhage diseases, and used them in the transfer learning procedures. Then, we experimented with various methods and techniques to create our classification model, including supervised learning with a ResNet backbone and transfer learning, self-supervised learning with MOCO, use of RNN models like LSTM to utilize the 3D context of CT data, vision transformers to obtain better fine-grained data features, and object detection techniques like YOLO and interesting findings were obtained.