Leveraging MobileNetV2 and deep learning innovation for high accuracy Plasmodium Vivax detection in blood smears

Malaria remains a significant public health challenge in regions where it is endemic. Pregnant women and young children are particularly vulnerable to the disease. Effective and timely diagnostic methods are crucial for reducing severe health outcomes. These methods help prevent deaths and lessen the widespread clinical and epidemiological burden on at-risk populations. However, traditional methods involved in the process of malaria parasite detections such as microscopic examination of blood smear by medical trained technicians is known to be time consuming, purely subjective and highly prone to errors. Therefore, Artificial Intelligence (AI) based OD (Object Detection) model like YOLO are preferred for overcoming these issues faced by traditional approaches as YOLO is known to be more rapid and precise by predicting bounding boxes and class probabilities than other models. However, existing YOLO model face challenges such as higher localization error, struggle with small objects and better accuracy of the model. Therefore, proposed research work focuses on employing YOLOv3 model with modified MobileNetv2 in backbone structure for classifying Plasmodium vivax (P. vivax) cells with the aim of improving the performance and speed of the model for detecting objects as MobileNetv2 is known for its faster processing and reduced resource consumption. However, accuracy is still measured as one of the key downsides for detecting and classifying the classes of thin blood smear, therefore modified MobileNetv2 is used, where proposed TCL (Transformed Convolution Layer) is employed at bottleneck layer, where weights are calculated based on different classes of image features thereby making the process more effective for classifying the infected and uninfected malaria cells of thin blood smear images effective. Besides, the performance of the proposed model is evaluated by implementing different metrics where the findings obtained are accuracy value of 1.00, precision value of 0.98, recall of 0.98, F1 score of 0.97 and mean average precision (mAP) value of 0.90. The major contribution of the study focuses on providing a better diagnostic approach for medical professionals in order to obtain improved results.