SINO-CT-Fusion-Net: A Lightweight Deep Learning Framework for Detection and Classification of Intracranial Hemorrhages

Deep learning has significantly improved medical diagnostics with its ability to learn the underlying complex patterns. A sinogram contains a sequence of X-ray projections of the patient into a lower dimensional space from different viewing angles, and a CT image is obtained as a result of applying reconstruction algorithms on the sinograms acquired by the scanner. While CT images are commonly used for automated diagnosis, recent developments have demonstrated that sinogram-based approaches can provide results on par with CT-based methods. This work leverages from the advantages of both approaches through the fusion of features learned from both those images. This paper presents a new lightweight deep learning model to detect and classify Intracranial Hemorrhages (ICH) through the fusion of high-level features learned from both sinogram and CT images. The proposed method is trained and evaluated on the publicly available RSNA ICH dataset. Furthermore, we analyze its multi-label classification capability in categorizing hemorrhages into five types. The proposed fusion model outperformed both CT-based and sinogram-based methods in general, it is particularly useful when there is less annotated training data and limited computational resources. The code and data can be found at https://github.com/sindhura234/SinoCTFusionNet