Application and comparison of object detection techniques to microscopic images of bacteria

Abstract

Microscopic interpretation of stained bacteria plays a vital role in diagnosing infectious diseases, including some prevalent bacterial infections like bacterial vaginosis (BV). Collected clinical samples are stained and inspected by experienced clinical laboratory technologists which is time-consuming and the results sometimes are operator-dependent. In the past decade, computer vision has been applied to reduce human workload and increase performance on repetitive tasks. In mid-2020, a Chinese research team published a convolutional neural network (CNN) model called NugentNet specialised in diagnosing BV using microscopic images of high vaginal swab (HVS) and yielded 82.4% sensitivity and 96.6% specificity at 5 images per second (Wang et. al, 2020). In this study, potential improvement to the modern artificial intelligence approach was explored. The introduction of a reliable and high-performance bacteria detection model may provide better BV diagnosis results, leading to higher sensitivity and specificity. Faster R-CNN and YOLOv5 architectures were the candidates for this experiment. Training time, inference time and mAP (mean average precision) were then compared to decide the most suitable detection model. Starting from healthy images with Lactobacillus spp., the practicability of each model architecture was demonstrated using Gram-positive rods at 1000x magnification. Subsequently, the object detection models were then trained on annotated HVS images with a mixture of mostly Lactobacillus spp. and Gardnerella spp. Diagnosis criteria were designed based on ordinal regression using training data, where all Gram-positive rods and Gram-negative coccus detected were treated as predictors. Using stratified random sampling with a train-test split of 3:1, classification accuracy, sensitivity and specificity of altered vaginal flora on the testing dataset were my main objectives. After obtaining the image-level diagnosis, similar ordinal regression techniques were repeated to average predicted bacteria counts to see if there will be further improvements. YOLOv5 small was found out to be the best performer with mAP@0.5 of 0.379, 3-class classification accuracy of 84.21% (case-level), sensitivity and specificity of altered vaginal flora of 86.67% and 86.96% respectively.