Simplistic refinement of self-supervised feature representations for classification of Brain strokes using contrastive learning on CT Images

The procedure for determining the cause of strokes makes heavy use of Computed Tomographic (CT) images. The existing methods try different Machine Learning (ML) based methods but haven't so far explored the self-supervised learning approach. In this study we have showcased a simple technique for improving the feature representation generated using Deep Learning (DL) models through self-supervised technique. We utilized (triplet) loss for generating different embeddings for normal and acute stroke cases and developed a simple strategy to improve the same for better differentiability. We showcase this refining technique, which improves discriminability and able to classify using simple ML model. The results are tested against various DL models for better generality with the best begin ResNet101 with 95.16% accuracy. We also compared the results based on different levels of refinement based on their architecture. Hence, the proposed system can be used in hospitals to analyze brain CT images.