An Optimized Transfer Learning Approach Integrat-ing Deep Convolutional Feature Extractors for Malaria Parasite Classification in Erythrocyte Microscopy

Kishor Kumar Reddy C¹Anisha P R¹Ahlam Almushharaf²Radhika Talla¹Jamel Baili³Yongwon Cho^4*Yunyoung Nam^4*

• ¹Department of Computer Science and Engineering, Stanley College of Engineering & Technology for Women, Hyderabad, India
• ²Department of Business Information Systems, Princess Nourah Bint Abdulrehman University, Saudi Arabia, Saudi Arabia, Saudi Arabia
• ³Department of Computer Engineering, College of Computer Science, King Khalid University, Abha, Saudi Arabia
• ⁴Department of Computer Science and Engineering, Soonchunhyang University, Asan, Republic of Korea

Background: Malaria, caused by Plasmodium parasites transmitted through bites from infected female Anoph-eles mosquitoes, results in severe symptoms such as anaemia and potential organ failure. The high prevalence of malaria necessitates reliable diagnostic methods to reduce the workload of microscopists, particularly in resource-limited settings. Methods: This paper evaluates the efficacy of an ensemble learning approach for automated malaria diagnosis. The proposed model integrates convolutional ensemble methods, combining outputs from transfer learning ar-chitectures such as VGG16, ResNet50V2, DenseNet201, and VGG19. Data augmentation and pre-processing techniques were applied to enhance robustness, and the ensemble approach was fine-tuned for optimal hy-perparameters. Results: The ensemble achieves a test accuracy of 97.93% by combining a evidence of CNN with multiple transfer learning models (VGG16, ResNet50V2, DenseNet201, and VGG19), with an F1-score and precision of 0.9793 each, outperforming standalone models like Custom CNN (accuracy: 97.20%, F1-score: 0.9720), VGG16 (accuracy: 97.65%, F1-score: 0.9765), and CNN-SVM (accuracy: 82.47%, F1-score: 0.8266). The method demonstrated effectiveness in classifying parasitized and uninfected blood smears with high reliability, ad-dressing the limitations of manual microscopy and standalone models Conclusion: The proposed ensemble learning approach highlights the potential of integrating transfer learning models to improve diagnostic accuracy for malaria detection. This scalable, automated solution reduces reliance on manual microscopy, making it highly applicable in resource-constrained settings and offering a significant advancement in malaria diagnostics.