JU-TBv1: a large-scale benchmark chest X-ray dataset for tuberculosis detection

Tuberculosis (TB) is a fatal lung disease caused by Mycobacterium tuberculosis. The early detection and diagnosis of TB is crucial as it becomes fatal if not diagnosed early. Artificial Intelligence, namely Deep Learning (DL) has shown promising results in precise and rapid detection of TB from various modalities such as chest X-rays, sputum smear images, patient clinical information, etc. However, DL techniques are data intensive and require large-scale datasets for efficient and accurate TB detection. This facilitates the need for a large-scale dataset for TB detection. In this study, we propose JU-TBv1, a large-scale open source publicly available dataset for TB detection. JU-TBv1 provides substantially larger scale and class-level duplicate-free TB dataset, with standardized benchmarking baselines. The dataset comprises of TB classification task from chest X-rays belonging to 2 classes, which are normal and TB positive. Furthermore this study uses a multi-stage filtering heuristic, TriCascade Algorithm for removing duplicate CXR images. The dataset has been used for training four popular Convolutional Neural Networks (CNNs) for TB detection, namely MobileNetV2, EfficientNetB0, ResNet50 and InceptioNetV3. The study also compares full-precision (fp32) and mixed-precision (fp16) training of the four CNN models for precision comparisons. The EfficientNet model outperforms the rest by achieving a test accuracy of 97.18% for fp32 and 97.15% for fp16. Additionally, four Transformer based architectures including ViT, Swin Transformer, LeViT and EVA-02 has been trained on JU-TBv1. The Transformer architectures show strong performance on the test set of JU-TBv1 with Swin Transformer outperforming the other architectures achieving an accuracy of 98.28% and F1-score of 0.98. Furthermore, to make the results more explainable, this study shows Grad-CAM analysis of the four CNN models for fp32 training.

Authors

Institutions

Publication Details

Journal
Advances in Computational Intelligence
Published
2026-11-14
DOI
https://doi.org/10.1007/s43674-026-00097-1
Primary Topic
COVID-19 diagnosis using AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

JU-TBv1: a large-scale benchmark chest X-ray dataset for tuberculosis detection

Debgandhar Ghosh, Pawan Kumar Singh
1 links
Advances in Computational Intelligence
COVID-19 diagnosis using AI
article

JU-TBv1: a large-scale benchmark chest X-ray dataset for tuberculosis detection

Debgandhar Ghosh, Pawan Kumar Singh
article en

Abstract

Tuberculosis (TB) is a fatal lung disease caused by Mycobacterium tuberculosis. The early detection and diagnosis of TB is crucial as it becomes fatal if not diagnosed early. Artificial Intelligence, namely Deep Learning (DL) has shown promising results in precise and rapid detection of TB from various modalities such as chest X-rays, sputum smear images, patient clinical information, etc. However, DL techniques are data intensive and require large-scale datasets for efficient and accurate TB detection. This facilitates the need for a large-scale dataset for TB detection. In this study, we propose JU-TBv1, a large-scale open source publicly available dataset for TB detection. JU-TBv1 provides substantially larger scale and class-level duplicate-free TB dataset, with standardized benchmarking baselines. The dataset comprises of TB classification task from chest X-rays belonging to 2 classes, which are normal and TB positive. Furthermore this study uses a multi-stage filtering heuristic, TriCascade Algorithm for removing duplicate CXR images. The dataset has been used for training four popular Convolutional Neural Networks (CNNs) for TB detection, namely MobileNetV2, EfficientNetB0, ResNet50 and InceptioNetV3. The study also compares full-precision (fp32) and mixed-precision (fp16) training of the four CNN models for precision comparisons. The EfficientNet model outperforms the rest by achieving a test accuracy of 97.18% for fp32 and 97.15% for fp16. Additionally, four Transformer based architectures including ViT, Swin Transformer, LeViT and EVA-02 has been trained on JU-TBv1. The Transformer architectures show strong performance on the test set of JU-TBv1 with Swin Transformer outperforming the other architectures achieving an accuracy of 98.28% and F1-score of 0.98. Furthermore, to make the results more explainable, this study shows Grad-CAM analysis of the four CNN models for fp32 training.

Advances in Computational IntelligenceVol. 6(2)
Jadavpur University (IN)
Good health and well-being
COVID-19 diagnosis using AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon