Malaria Parasite Classification in Thin Blood Smears Based on CNN with SpinalNet and GAN

Abstract

This research presents a novel framework integrating Generative Adversarial Networks (GANs) with a parameter-efficient Convolutional Neural Network (CNN) to automate malaria parasite classification from thin blood smear microscopy images. A major bottleneck in automated malaria diagnosis is severe clinical data imbalance, where minority species (e.g., Plasmodium ovale, P. malariae) are heavily underrepresented, causing standard deep learning models to default to majority-class predictions. To address this, our methodology employs a Wasserstein GAN with Gradient Penalty (WGAN-GP) to synthesize high-fidelity microscopic images, effectively balancing the training distribution prior to classification. Following rigorous preprocessing---including Contrast Limited Adaptive Histogram Equalization (CLAHE) and aggressive data augmentation to mitigate staining variability---the balanced dataset is processed by a CNN for feature extraction. Crucially, traditional fully connected layers are replaced with a SpinalNet classification head. This architectural innovation processes extracted features sequentially, which drastically reduces the number of trainable dense parameters while acting as a structural regularizer against overfitting on the synthetic data. Evaluated on strictly real, hold-out clinical test sets, the proposed GAN-CNN-SpinalNet pipeline demonstrates a strong capability to identify rare parasite species without sacrificing performance on the P. falciparum majority class. Assessed across standard metrics, including accuracy and Macro F1-score, this approach successfully overcomes data scarcity to deliver a robust, and computationally efficient diagnostic tool for resource-limited settings.