Automatic classification of microscopic images of fecal samples based on machine learning techniques to diagnose cases of Giardia parasite

Abstract

Giardiasis is a significant cause of diarrheal disease and a major public health concern. The manual analysis of stool samples is time-consuming and requires a high degree of expertise, which has motivated the development of computer-aided diagnosis systems to automate the classification of giardia subtypes from stool microscopic images. In this study, we propose a novel deep learning-based approach for the automatic classification of giardia subtypes from stool microscopic images. Specifically, we used a dataset of 1,934 images taken from 48 slides, where 40 slides contained 1,610 unique images for training and evaluating the network with an 80/20 split, and 8 slides contained 324 unique images from three different classes for testing the network. The number of unique samples for training and evaluating the network in each class were 570 normal, 410 cyst, and 630 trophozoite samples. The test dataset included 92 cyst samples, 113 normal samples, and 119 trophozoite samples.

We employed four deep learning models for image classification: VGG16, ResNet50, Xception, and EfficientNet-B0. Prior to training, the images were preprocessed using non-local means denoising and were transformed from RGB to grayscale. Local histogram equalization was then applied to the images to enhance their contrast. The models were trained with dropout technique to prevent overfitting. The input images were of size 512 x 512 with 3 channels. Our proposed automated method for the classification of Giardia lamblia parasite into three categories including normal, cyst, and trophozoite, based on stool samples, using pre-trained deep learning models, achieved promising results. The results of the test dataset showed that EfficientNetB0 and ResNet50 models achieved the best performance with accuracy of 96.94% and 96.02%, respectively. These findings suggest that our approach has the potential to provide a practical and effective solution for the automated classification of giardia subtypes, which could aid in the diagnosis and treatment of giardiasis.