Towards Reliable Malaria Diagnosis: Hybrid CNN Framework Based on VGG16 with Data Augmentation and Dropout-Enhanced Training Strategy
Md. Tofael Ahmed Bhuiyan, Md. Nazim Uddin, Shahriar Manzoor, Khandaker Mohammad Mohi Uddin · 2025
Plasmodium parasites produce malaria, which may be very dangerous if left untreated. While computer-aided procedures allow for quicker and more accurate detection, conventional diagnostic methods are often expensive and subject to human mistake. Deep learning works especially well for classifying malaria because it makes use of a wealth of visual data. This study employs transfer learning with pre-trained convolutional neural networks (CNNs), such as VGG16, ResNet50, and EfficientNetB3, to differentiate between images of cells infected with malaria and those that are not. On a well-known malaria dataset, VGG16 has the best classification accuracy of 98.15 % among them, together with outstanding specificity and sensitivity. To maximize model performance, the approach combines data augmentation with a customized CNN architecture built on the VGG16 architecture, which is improved by dropout regularization along with training callbacks. Numerous tests confirm the efficacy of the suggested approach, offering a dependable and automated way to diagnose malaria.