Computer Sciences and Information Technology
Title : | Computationally Lightweight Convolutional Neural Network for Generalizable Chest X-ray Diagnosis |
Area of research : | Computer Sciences and Information Technology |
Principal Investigator : | Dr. Angshuman Paul, Indian Institute Of Technology Jodhpur (IITJ), Rajasthan |
Timeline Start Year : | 2022 |
Timeline End Year : | 2024 |
Contact info : | angshuman.bankura@gmail.com |
Equipments : | GPU Workstation |
Details
Executive Summary : | Chest x-rays are widely used imaging tests due to their ability to detect thoracic abnormalities and are cost-effective and fast. Deep learning (DL) approaches have revolutionized automated medical image analysis, including chest x-ray analysis, by nearly matching human performance in clinical tasks. These methods can act as decision support systems and improve the diagnosis of human experts, particularly in Indian rural healthcare scenarios. In primary health centers, primary physicians may not have access to radiologists, which could lead to inaccurate diagnosis of chest x-ray abnormalities. An automated decision support system could help primary physicians diagnose abnormalities accurately, prioritize severe cases for referral to advanced facilities, and treat milder cases more effectively. This would result in quicker treatment start times and better prognosis. However, two major challenges may hinder the applicability of such a system in Indian rural health centers: computationally intensive DL models that require expensive GPUs and are not generalizable across different datasets. To address these issues, a generalizable computationally lightweight convolutional neural network (CNN) model is proposed. The model will focus on high reproducibility of results and use publicly available large scale chest x-ray datasets for experiments. Cross-dataset experiments will be conducted to test the model's generalizability, with a particular focus on designing a classifier for multi-label diagnosis of chest x-rays. This model with high accuracy, reproducibility, and generalizability would be clinically useful as a decision support system, especially for primary physicians in primary health centers. Its computational lightweight nature makes it cost-effective and deployable in rural health centers. |
Total Budget (INR): | 28,72,820 |
Organizations involved