Developing a Leukemia Diagnostic System Based on Hybrid and Ensemble Deep Learning Architectures

This research designed and implemented a leukemia diagnostic system targeted for a real clinical environment based on deep learning approaches. The dataset consists of 15,135 total images. This research develops four independent models (VGG19 1 epoch, VGG19 30 epochs, ResNet50 1 epoch, and ResNet50 30 epochs), two Hybrid models (trained at 1 epoch and 30 epochs), and four Ensemble models (two Ensemble models of VGG19 and ResNet50 and two Ensemble models with an additional Hybrid model). All models are pre-trained on ImageNet. By using transfer learning, the models were fine-tuned (further trained) on the leukemia domain at a much greater speed as the existing layers will have benefitted from the pre-training done on ImageNet. This research indicates that Hybrid models can help improve predictive capabilities by leveraging different feature patterns extracted from running images through two different architectures. Meanwhile, Ensemble models will take the prediction votes from multiple final model outputs to further incorporate different model capabilities and also help generalize. Among all independent models, the best model is ResNet50 30 epochs, which achieved an accuracy of 84%. Among the Hybrid models, the best model is Hybrid 30 epochs, which achieved an accuracy of 84%. Ensemble 4 (Hybrid 30 epochs, VGG19 1 epoch, and ResNet50 1 epoch) achieved an accuracy of 86%, which is 2% better than the second-best model, Ensemble 2. The diagnostic system developed in this research can be used in other medical diagnostic applications.