Hyperparameter Optimization of DenseNet-121 for Beta Thalassemia Major Detection on Android

Deep learning technology has made significant strides in diagnosing various diseases. One of the most crucial factors is selecting the appropriate hyperparameter combination, which can enhance the accuracy of medical diagnosis. However, choosing the best set of hyperparameters among the possible combinations can be a significant challenge. This study focuses on optimizing the hyperparameters of the DenseNet-121 architecture using a random search for the detection of Beta Thalassemia Major. The results are applied to an Android app for beta thalassemia major detection. The training process involves 14 trains to optimize batch size and learning rate. The dataset consists of 506 facial images divided into two classes. The proposed DenseNet-121 model utilizes a batch size of 8 and a learning rate 0.001. The results indicate an accuracy of 97% and a validation accuracy of 93 %, with corresponding losses of 0.1 % and 0.2%. During testing, the accuracy for healthy and Thalassemia images ranged from 84 % to 99%. These findings underscore the approach's effectiveness in identifying potential cases of Beta Thalassemia Major.