Performance Comparison of a Hybrid Convolutional Neural Network-Long Short-Term Memory and CNN Model for Malaria Diagnosis using Microscopic Blood Smears

Abstract

Aim: This study evaluates and compares the performance of a lightweight Convolutional Neural Network (CNN) and a hybrid CNN–Long Short-Term Memory (CNN–LSTM) model for classifying malaria-infected and uninfected microscopic blood smear images, with emphasis on performance–efficiency trade-offs. Methods: A dataset of 27,558 images was preprocessed through normalization and resizing, then split into training, validation, and test sets (60:20:20). Both models were implemented using TensorFlow and Keras. The CNN comprised three convolutional layers, while the CNN–LSTM incorporated sequential learning with significantly fewer parameters. Performance was evaluated using accuracy, precision, recall, F1-score, confusion matrices, and ROC curves with AUC. Results: The CNN achieved 94.90% accuracy (AUC: 94.91%), demonstrating strong performance despite its simplicity. The CNN–LSTM, with a much smaller parameter size, achieved 91.42% accuracy and a higher AUC of 97.15% (95% CI: 0.97–0.98). Comparative analysis with models such as VGG16, VGG19, ResNet variants, MobileNetV2, Xception, InceptionV3, and DenseNet201 showed that the CNN performs competitively with lower computational cost. Conclusion: Lightweight CNN architectures can deliver performance comparable to deeper models, while CNN–LSTM offers a compact alternative with strong class separability, supporting efficient and scalable malaria diagnosis in resource-constrained settings.