Authors (including presenting author) :
WONG RCW (1), CHAN RKF (2), NGAN RHH (2), WAN J (3), WONG J (3), LUI K (3), TAO JHY (1), WONG DSH (1), Chow VCY (1)
Affiliation :
(1) Department of Microbiology, PWH, (2) ITD, NTEC, (3) IT&HI, HAHO
Introduction :
Millions of people have suffered from intestinal parasitic infections worldwide. Laboratory diagnosis of helminth eggs (ova) in faecal specimen is conventionally based on microscopic examination of concentrated wet mount prepared from the formalin-ethyl acetate sedimentation. Morphological features and size dimension determined by the calibrated ocular micrometer of microscope are both crucial in the identification of ova. Medical Laboratory Technologist (MLT) will examine microscopically, if ova is suspected, then MLT will validate the observation with reference to the literature. However, such process is prone to human errors and rely on the competency of MLT. Hong Kong is lacking of local positive samples, and hence inadequate training materials to allow MLTs to gain adequate expertise. To resolve such problem, an automatic recognition program would be desirable.
Objectives :
We aim to apply AI technology in the identification of ova from faecal specimens and provides a user-friendly tool to assist MLTs.
Methodology :
Ova was detected by a convolutional neural network-based “You Only Look Once” (YOLOv5) object detection deep-learning model. Image dataset was collected from clinical, EQAP samples, IEEE open dataset and internet with manual validation by well-experienced MLTs. It included 5373 images from 14 parasites included nematodes (i.e. Enterobius vermicularis, Ascaris lumbricoides, Trichuris trichiura, Capillaria philippinensis & hookworm), cestodes (Taenia species, Hymenolepis nana, Hymenolepis diminuta & Diphyllobothrium latum) and trematodes (i.e. Schistosoma mansoni, Schistosoma japonicum, Clonorchis sinensis, Paragonimus species & Fasciolopsis buski). The preprocessing of images included random rotation, stretches, warps, insert noise, brightness & contrast adjustment for reality simulation and generate extra data to enhance the size of dataset. The deep learning process utilized the dataset in the following ratio: 72%, 18% & 10% for the training, validation, and testing sets, respectively. To include the size parameter in the result interpretation, users are required to input the dimension according to their own calibrated micrometer. The size reference range was built in the program for each of the species according to the information published by the CDC.
Result & Outcome :
This model showed satisfactory performance metrics with 0.97 mAP@0.5, accuracy of 0.98 and F1-score of 0.96 among the testing sets. Here is the first milestone to apply AI technology in clinical microbiology laboratories in Hong Kong. It serves act as a confirmatory tool in the identification of ova and local training database with plenty of images as users can upload their microscopic images. It also reduces the time required for identification (from image upload to answer will take ~1 min) compare with manual validation with literature and provides reproducible results. The number of collected images will growth continuously to improve the object detection accuracy and coverage of dataset will expand to include other species. Our ultimate goal is to develop an automatic examination system and provides a total solution for MLT in the examination of ova by screening out the negative samples, while sorting the positive one for manual validation. However, it will require implementation of integrated microscopy system (~HK$0.6M) that can automate digital microscopic slide scanning, image capturing and development of IT interface for uploading the images to the AI Portal is also necessary.
