Faster convolutional neural network inferencing for screening red blood cell disease

Abstract

The peripheral blood smear is a blood test, in which its purpose is to detect or confirm diseases in patients by identifying the morphological characteristic of the blood. The blood smear could only be interpreted through microscope by a skilled laboratory physician or haematologist which is time consuming and would result in other complication such as higher labour cost and error rate. In Malaysia particularly, the blood smear analysis is only performed in certain general hospital, which have a hematopathology unit, thus having more test to perform at once and ultimately further increases time. Hence, the approach taken to address this issue is by creating an automated blood smear screening tool. Typically feature extraction was done manually in computer vision to identify the disease but this method might introduce a drop in accuracy if the design was under or over segmented. Therefore, this research proposes a convolutional neural network (CNN) to screen diseases from the blood smear images. The focus would be on developing an optimized CNN model in terms of inferencing speed for faster object detection of sickle cell disease. Adopting a popular object detection CNN algorithm i.e., YOLOv3 which has a high inferencing speed, this research performed transfer learning, modify certain CNN attributes by adding a residual block and applying Gaussian filter prior to model training on blood smear images to optimize speed and accuracy. The proposed model, which is to train YOLOv3 with Gaussian filtered dataset achieved an inferencing speed of 379ms per image with a mean average precision of 72.74%. The algorithm had a fast-inferencing speed but suffers in term of accuracy due to insufficient dataset and failing to detect overlapping cells. To solve this, data augmentation technique and sharpening edge of the cell should be performed. In short, this project contributed towards existing literature by developing CNN that prioritize inferencing speed and implement YOLOv3 object detection for localization and classification of sickle cell disease.