Comparative Evaluation of Custom Convolutional Neural Networks and Efficientnet-B3 for Malaria Cell Image Classification: Impact of Targeted Data Augmentation on Model Performance



Mohit R ; Milani E ; Amini A ; Ahani M ; Nazari E ; Aldaghi T
Source: Iranian Journal of Blood and Cancer Published:2025

Abstract

Background: Malaria diagnosis with thin blood smears remains labor-intensive and relies on the operator. Deep learning could enable accurate automation. Objective: Compare four convolutional approaches for classifying parasitized versus uninfected erythrocytes and to evaluate whether targeted image-quality augmentations enhance performance. Materials and Methods: We used the balanced NIH/Kaggle dataset, which included 13,780 parasitized and 13,780 uninfected samples. Data were split stratified into training, validation, and test sets (70/15/15). Images were resized to 256×256 and normalized. Four experiments were conducted: (1) a custom CNN; (2) the same CNN with targeted augmentation applied to 20% of training samples per class—using Contrast Limited Adaptive Histogram Equalization [CLAHE] and controlled brightness adjustment—and augmented images were added back to the training set (totaling 30,864 images); (3) a soft-attention parallel CNN (SPCNN); and (4) transfer learning with EfficientNet-B3 on 300×300 inputs with full fine-tuning. Evaluation metrics included accuracy, precision, recall, F1 score, and AUC-ROC. Results: EfficientNet-B3 achieved the highest performance with a validation accuracy of 0.9741, 98% precision, 96% recall, an F1 score of 0.97, and an AUC-ROC of 0.9964. SPCNN was competitive but slightly lower, with a validation accuracy of 0.9652, 98% precision, 95% recall, an F1 score of 0.96, and an AUC-ROC of 0.9909. The baseline CNN had a validation accuracy of 0.9649, 97% precision, 94% recall, an F1 score of 0.96, and an AUC-ROC of 0.9910. Targeted augmentation resulted in negligible change compared to the baseline CNN, with a validation accuracy of 0.9647, an F1 score of 0.96, and an AUC-ROC of 0.9908, indicating limited added discriminative value for this dataset. Conclusion: EfficientNet-B3 outperformed SPCNN and custom CNNs. The CLAHE/brightness strategy applied to 20% of training images and added back to the dataset did not significantly improve generalization. External validation and prospective field testing are necessary before clinical deployment. © 2025 Elsevier B.V., All rights reserved.