Hybrid Deep Learning and Machine Learning Approach for Sickle Cell Disease Diagnosis Using ResNet and Random Forest Classifier

This study presents a state-of-the-art deep learning classifier based on the residual network architecture developed for diagnosing sickle cell disease (SCD). The technique is a combination of deep learning method for extraction and conventional machine learning approach for classifiers using pre-trained model ResNet50. The dataset is pre-processed (resizing the images to 224x224 pixels, converting them into greyscale, and then creating three-channel formats to save it in a form that could be fitted with ResNet50). This additional data is generated using diversatilation strategies like horizontal flip, rotation, zoom etc which helps in better model generalization. In this paper, we replace the softmax layer of deep features extracted by ResNet50 model with Random Forest classifier to improve classification accuracy. The model is tested on SCD dataset achieving an excellent accuracy of 96.53% with a great performance across classes, i.e., precision rate for circular, elongated and other being 94.49%, 97.62% and 97.54% respectively, which are quite impressive metrics considering the distribution of classes in comparison to number of input samples per class. The F1 scores also demonstrate a good balance between precision and recall. According to ROC analysis, this CDCNN showed a very promising discrimination capability of 0.97 AUC in case of both circular and elongated classes, the remaining class demonstrated bog highest value of 0.98. We then present benchmarking results that portray strengths of the new suggested model with respect to current classifiers and thus potential for highly accurated predictions in an application targeted at medical diagnostics.