Title : Automated detection of normal-pressure hydrocephalus using DINOv2 vision transformer and DETR detection transformer technologies on CT imaging
Abstract:
Background: Normal Pressure Hydrocephalus (NPH) is a neurological condition characterized by the enlargement of the lateral ventricles in the brain despite normal cerebrospinal fluid pressure. Although this condition is treatable, it is often underdiagnosed due to its similarity to other conditions that resemble it. NPH is differentiated from other similar conditions using the pattern by which the lateral ventricles enlarge, which can be determined by radiological markers such as the Evans Index and Callosal Angle. These measurements are calculated using images of the brain, primarily CT scans. Although these measurements provide the data needed to diagnose NPH, they are not consistently calculated during routine clinical evaluation, contributing to delayed or missed diagnosis. This study aims to create a Machine Learning based automated tool that is able to accurately calculate these radiological markers, and assist primary care physicians in diagnosing NPH as efficiently and accurately as possible.
Methods: This study developed a transformer-based machine learning platform to automatically detect NPH-associated radiological markers from CT imaging. 1,000 CT slices (500 axial and 500 coronal) from open-access datasets were preprocessed, and the coronal images were manually annotated to train a DINOv2 Vision Transformer and Detection Transformer (DETR) to identify and segment the lateral ventricles. Using the detected ventricular boundaries, code was then developed to teach the model to automatically calculate the Evans Index and Callosal Angle, converting imaging data into quantitative diagnostic measurements. To evaluate performance, 88 axial and 88 coronal CT images ranging from normal to hydrocephalus were independently measured by a radiologist to calculate Evans Index and Callosal Angle using established clinical thresholds. The same images were then processed through the ML system, which automatically calculated these measurements and generated NPH predictions. The model’s results were compared with the radiologist’s measurements, and accuracy, sensitivity, specificity, false positive rate, and false negative rate were calculated to assess model performance.
Results: The model demonstrated strong diagnostic performance, achieving 92.05% accuracy using Evans Index measurements and 93.18% accuracy using Callosal Angle measurements, with AUC values of 0.965 and 0.896, respectively. These results showed strong agreement with manual neuroradiologist measurements, indicating that transformer-based models can reliably quantify radiological markers associated with hydrocephalus.
Conclusion: This project represents a novel application of Vision Transformers and Detection Transformers to calculate quantitative radiological measurements directly from CT imaging, an approach that has not previously been used to automate NPH marker detection. By enabling automated identification of critical diagnostic markers, this system has the potential to support physicians in recognizing normal-pressure hydrocephalus earlier and more consistently. Not only can this be a viable tool that primary care physicians use to speed the diagnosis process, but also be a publicly available tool that can be used free of cost.
