Speaker
Discussion (methods + interpretation)
We utilized a public dataset of 4,200 chest X-rays (3,500 Normal, 700 TB-positive). To address class imbalance and prevent overfitting, data augmentation techniques were applied alongside standard image resizing. We implemented transfer learning using pre-trained convolutional neural networks (ResNet-50/MobileNetV2) by freezing base layers and fine-tuning final classification layers. Model performance was evaluated prioritizing sensitivity over accuracy. Crucially, Grad-CAM visual heatmaps were integrated to highlight abnormal lung regions, resolving the "black box" issue and offering clinicians interpretable visual rationale to validate AI predictions confidently.
Aim & Purpose
This study aims to develop an AI-driven automated screening system using deep learning to triage chest X-rays in real time. By providing instant risk classification, it enables early isolation of contagious TB patients and prioritizes urgent cases for doctor review, mitigating diagnostic bottlenecks in public healthcare.
Background
Pakistan faces a severe Tuberculosis burden, with over 140 daily deaths. Chest X-rays are the primary diagnostic tool, but severe shortages of radiologists in overcrowded public and rural hospitals cause critical diagnostic delays. During these waiting periods, contagious patients risk spreading infection. An automated screening tool is urgently needed to provide immediate triage and flag high-risk cases.
Result
The fine-tuned transfer learning model achieved high classification performance, significantly outperforming standard CNNs trained from scratch. By prioritizing recall/sensitivity over baseline accuracy, the model effectively minimized false negatives—crucial for medical screening. Furthermore, Grad-CAM heatmaps accurately localized lung lesions, providing doctors with real-time, interpretable visual evidence for rapid triage
| Email address | 24G-BCS009@agu.edu.pk |
|---|---|
| Phone number | 03322107506 |