Development of Self-Supervised Learning with Dinov2-Distilled Models for Parasite Classification in Screening
Date
2023
ISBN
979-835030446-6
Journal Title
Journal ISSN
Volume Title
Resource Type
Conference paper
Publisher
Institute of Electrical and Electronics Engineers Inc.
Journal Title
Development of Self-Supervised Learning with Dinov2-Distilled Models for Parasite Classification in Screening
Recommended by
Abstract
At present, parasitic infections in humans, such as intestinal parasitic infections and soil-transmitted helminth (STH) infection, remain a public health concern, with screening methods that are simple but time-consuming and require parasitology experts. Microscopy images are increasingly being used to aid diagnosis but creating labels for supervised learning (SL) is a time-consuming, labor-intensive, and costly process. Self-supervised learning (SSL) is a deep learning approach that aims to train models to represent features in unlabeled datasets using automatically generated labels or annotations from the data itself, rather than explicitly labeled human-labeled labels. It is an appropriate method to address the challenges associated with the difficulty of labeling large datasets. A pretrained model that has learned useful data representations from an SSL task is fine-tuned using labeled data to perform well on a specific downstream task. DINOv2 is an SSL model based on the Vision Transformer (ViT) architecture. In this study, we aim to create a model for screening for helminth egg infection using a fine-tuned Dinov2 with a classification layer head to demonstrate that dataset sizes of 1% and 10% are sufficient when compared to SL model. Rather than SL, which requires a significant amount of human data labeling and is generally impractical, the model developed in this study is expected to be used in active surveillance in the future. © 2023 IEEE.
Description
Citation
2023 15th International Conference on Information Technology and Electrical Engineering, ICITEE 2023