Answers are grounded in this paper and cite their sources.
Malaria remains a major public health challenge, particularly in endemic regions with limited access to trained microscopists and laboratory infrastructure. Conventional microscopy is labor-intensive, expertise-dependent, and affected by observer fatigue and inter-observer variability. In these settings, automated support is more likely to be trusted and used safely when microscopists can check the basis of each prediction. This study presents an explainable deep-learning benchmark for automated malaria cell classification using the NIH Malaria Cell Images Dataset, comprising 27,558 thin blood smear images. Five convolutional neural network backbones (EfficientNetV2-S, VGG16, VGG19, ResNet34, and ResNet50) were compared, and MixUp and CutMix augmentation were examined for their effects on generalization. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to show which cell regions drove each prediction, so that a reviewer can judge whether a positive call is supported by parasite-like structures or by staining artifacts. Models were evaluated with 5-fold stratified cross-validation using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (ROC-AUC), together with learning curve analysis. ROC-AUC was the primary metric for the Friedman test with Nemenyi post-hoc analysis. EfficientNetV2-S achieved its best performance under the CutMix scenario, with a mean accuracy of 97.56 ± 0.30%, ROC-AUC of 99.72 ± 0.06%, and F1-score of 97.53 ± 0.31%. The Friedman test showed significant differences among backbones in all augmentation scenarios, also after Holm correction, although the absolute ROC-AUC differences were below 0.002. Nemenyi post-hoc analysis showed the significant superiority of EfficientNetV2-S over ResNet34 in all scenarios and over VGG19 under CutMix, but not over every baseline. Grad-CAM maps for cases selected by a fixed rule suggested that EfficientNetV2-S focused on parasite-like stained inclusions in correctly classified infected cells, whereas misclassified cells had probabilities close to the decision threshold and activations on faint staining irregularities or cell boundaries. A deletion-insertion check on 200 validation images supported the faithfulness of these maps for parasitized predictions but not for uninfected predictions. These observations were not validated by experts, but such heatmaps could help a microscopist confirm or reject a flagged cell. EfficientNetV2-S had the fewest parameters (20.84 million), the lowest number of multiply-accumulate operations (3.54 GMACs), and the smallest model size (79.49 MB), but was slower on GPU than all four baselines (23.87 vs. 6.12-10.80 ms per image), so its compactness refers to size rather than speed. Overall, EfficientNetV2-S is a promising backbone for automated malaria cell classification, offering strong discriminative performance, a compact model size, and qualitative Grad-CAM observations that may support human-in-the-loop screening. External validation on clinically diverse datasets and expert assessment of explainability outputs are required before clinical deployment.
No takes yet. Share an insight, caveat, or question.
Siswantomo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: