Explainable and Quantized Neural Networks for Generalisable Malaria Cell Classification

Malaria affects children and adults and caused an estimated 610,000 deaths worldwide in 2024, with approximately 95% occurring in the African region. Microscopic identification of erythrocytes infected by Plasmodium is important for malaria diagnosis but can be affected by workload, fatigue, limited expertise and variations in blood-smear preparation. Although previous neural-network classifiers have reported high accuracy on malaria-cell images, random image-level splitting may place cells from the same patient in both training and testing sets. This patient overlap compromises the independence of the evaluation and may produce an optimistic estimate of performance on previously unseen patients. This study evaluated a custom convolutional neural network, MobileNetV3Small, EfficientNetB0 and ResNet50 using the NIH malaria dataset containing 27,558 balanced images of parasitized and uninfected erythrocytes. Patient identifiers were used to form non-overlapping training, validation and held-out test groups. Model evaluation combined sensitivity, specificity, F1 score, ROC-AUC, PR-AUC and confusion-matrix analysis with parameter count, model size, inference latency and throughput. Grad-CAM was used to examine spatial evidence in correct and incorrect predictions. The selected model was converted to TensorFlow Lite FP32, FP16 and INT8 formats to assess quantization-performance trade-offs. MobileNetV3Small achieved the best performance-efficiency balance, with 96.47% accuracy, 95.69% sensitivity, 97.24% specificity and ROC-AUC of 0.9930. FP16 quantization reduced model size from 3.59 MB to 1.87 MB while retaining 96.44% accuracy. INT8 produced a smaller model but reduced accuracy to 93.54% and specificity to 90.32%. The findings show that FP16 MobileNetV3Small offers a favourable balance of classification performance, interpretability and deployment efficiency. The proposed framework classifies isolated erythrocytes as parasitized or uninfected, providing a foundation for future whole-slide analysis, Plasmodium species identification and parasitaemia estimation.