Malaria Parasite Detection in Thick Blood Smear Microscopic Images Using Modified YOLOV3 and YOLOV4 Models

Abstract

Background Information: Manual microscopic examination is still the ”golden standard” for malaria diagnosis. The challenge in the manual microscopy is the fact that its accuracy, consistency and speed of diagnosis depends on the skill of the laboratory technician. It is difficult to get highly skilled laboratory technicians in the remote areas of developing countries. In order to alleviate this problem, in this paper, we propose and investigate the state-of-the-art one-stage and two-stage object detection algorithms for automated malaria parasite screening from thick blood slides. Methods: YOLOV3 and YOLOV4 are state-of-the-art object detectors both in terms of accuracy and speed; however, they are not optimized for the detection of small objects such as malaria parasite in microscopic images. To deal with these challenges, we have modified YOLOV3 and YOLOV4 models by increasing the feature scale and by adding more detection layers, without notably decreasing their detection speed. We have proposed one modified YOLOV4 model, called YOLOV4-MOD and two modified models for YOLOV3, which are called YOLOV3-MOD1 and YOLOV3-MOD2. In addition, we have generated new anchor box scales and sizes by using the K-means clustering algorithm to exploit small object detection learning ability of the models. Results: The proposed modified YOLOV3 and YOLOV4 algorithms are evaluated on publicly available malaria dataset and achieve state-of-the-art accuracy by exceeding the performance of their original versions, Faster R-CNN and SSD in terms of mean average precision (mAP), recall, precision, F1 score, and average IOU. For 608 x 608 input resolution YOLOV4-MOD achieves the best detection performance among all the other models with mAP of 96.32%. For the same input resolution YOLOV3-MOD2 and YOLOV3-MOD1 achieved mAP of 96.14% and 95.46% respectively. Conclusions: Th experimental results demonstrate that the performance of the proposed modified YOLOV3 and YOLOV4 models are reliable to be applied for detection of malaria parasite from images that can be captured by smartphone camera over the microscope eyepiece. The proposed system can be easily deployed in low-resource setting and it can save lives.

Malaria Blood Smear Classification Using Deep Learning and Best Features Selection

“…From this table, it is demonstrated that the proposed method outperforms the existing techniques in terms of accuracy, training time, and prediction speed. Reference Accuracy (%) [36] 97.00 [19] 97.30 [16] 97.47 [17] 97.70 Ours 99.67…”

“…Several studies have been proposed for malaria blood smear detection, segmentation, and classification [16,17]. Quan et al [18] utilized 27,558 malaria blood smear images from the National Institute of Health (NIH) and expanded wielding rotation, zooming, and flipping.…”

Tile-Based Microscopic Image Processing For Malaria Screening Using Deep Learning Approach

“…Another study by [37] proposes an ensemble of pretrained and custom CNN models for classification of infected and uninfected RBC cells segmented from thin blood smear microscopic images. In [38], [39] modified versions of YoloV3 and YoloV4 models are proposed to improve P. falciparum detection capability of SOTA deep learning models on thick smear microscopic images and to make detection models lightweight to be integrated with mobile phone-based diagnosis application.…”

“…In this study, three different YOLOV4 based object detection models were evaluated to identify the best model with an optimal trade-off between detection performance and inference speed. The first detection network, which is called YOLOV4-MOD, is based on previous work [39] which is modified to improve small object detection performance of the original YOLOV4 model with minimal computation cost. This model has a large number of trainable parameters and more convolution layers.…”

“…Among these lightweight models, the first model, which is called YOLOV4-tiny, has two detection heads while the second detection model, which is called YOLOV4-tiny-3l, has three detection heads. The addition of more detection heads on YOLO-based models enabled better detection performance for small objects [39]. These two lightweight networks have a small number of trainable parameters and fewer convolution and max-pooling layers compared to the large size model.…”