Document Type
Thesis
Date of Award
2023
Degree Name
Master of Science (MS)
Department
Computer Science
First Advisor
KC Santosh
Abstract
Deep Learning (DL) has an extensively rich state-of-the-art literature in medical imaging analysis. However, it requires large amount of data to begin training. This limits its usage in tackling future epidemics, as one might need to wait for months and even years to collect fully annotated data, raising a fundamental question: is it possible to deploy AI-driven tool earlier in epidemics to mass screen the infected cases? For such a context, human/Expert in the loop Machine Learning (ML), or Active Learning (AL), becomes imperative enabling machines to commence learning from the first day with minimum available labeled dataset. In an unsupervised learning, we develop pretrained DL models that autonomously refine themselves through iterative learning, with human experts intervening only when the model misclassifies and for a limited amount of data. We introduce a new terminology for this process, calling it mentoring. We validated this concept in the context of Covid-19 in three distinct datasets: Chest X-rays, Computed Tomography (CT) scans, and cough sounds, each consisting of 1364, 4714, and 10,000 images, respectively. The framework classifies the deep features of the data into two clusters (0/1: Covid-19/non-Covid-19). Our main goal is to strongly emphasize the potential use of AL in predicting diseases during future epidemics. With this framework, we achieved the AUC scores of 0.76, 0.99, and 0.94 on cough sound, Chest X-rays, and CT scans dataset using only 40%, 33%, and 30% of the annotated dataset, respectively. For reproducibility, the link of implementation is provided: https://github.com/2ailab/Active-Learning.
Subject Categories
Computer Sciences
Keywords
Active Learning, Covid-19, Epidemics, Human-in-the-loop Machine Learning, Mentoring
Number of Pages
63
Publisher
University of South Dakota
Recommended Citation
Nakarmi, Suprim, "MENTORING DEEP LEARNING MODELS FOR MASS SCREENING WITH LIMITED DATA" (2023). Dissertations and Theses. 184.
https://red.library.usd.edu/diss-thesis/184