Hyperbolic Residual Support Vector Machine Model for Sickle Cell Disease Classification

This work proposed a hybrid model of hyperbolic embedding with transfer learning using ResNet50 and support vector machine (SVM) in this research work for classifying medical image data to Sickle Cell Disease (SCD) It is used in hyperbolic geometry to improve feature representation by capturing inherent hierarchical relations of the data better. This study uses a pre-trained ResNet50 model to extract features, and then fine-tune the model to classify SCD. An SVM classifier then classifies the extracted features. Overall prediction accuracy was 96%, with per class performance statistics as: Circular [Precision: 0.97; Recall: 0.95; F1-score: 0.96]; Elongated [Precision: 1.00; Recall: 0.96; F1-score: 0.98]; Other - [Precision: 0.90; Recall: 0.97]. Model has also used GradCAM XAI to visualize the affected blood cell in image. In the benchmarking, our model has achieved 2.69% better mean accuracy to decision tree based model.