JU-TBv1: A Large-scale Benchmark Chest X-Ray Dataset for Tuberculosis Detection

Abstract

JU-TB is a large-scale, single-modality chest X-ray (CXR) dataset curated for binary tuberculosis (TB) detection, designed to support robust benchmarking of deep learning models. The dataset contains 55,336 high-quality CXR images categorized into TB-positive and Normal classes, making it one of the largest publicly available TB CXR datasets. To ensure data integrity and prevent leakage, JU-TB employs a novel TriCascade duplicate removal algorithm, which combines perceptual hashing, structural similarity (SSIM), and pixel-wise verification to eliminate duplicate and near-duplicate images sourced from multiple open datasets. This guarantees that training, validation, and test splits are strictly non-overlapping. The dataset follows a standardized split: 1. Training: 47,035 images 2. Validation: 5,533 images 3. Test: 2,768 images JU-TB is specifically curated to be DL-ready, enabling training of data-hungry architectures such as CNNs and Transformers. Baseline benchmarking on MobileNetV2, EfficientNetB0, ResNet50, and InceptionNetV3 demonstrates strong performance, with EfficientNetB0 achieving over 97% accuracy, validating the dataset’s reliability. **Key Features** - Large-scale TB vs Normal CXR dataset - Strict duplicate and near-duplicate removal - Leakage-free train/val/test splits - Suitable for benchmarking, transfer learning, and explainable AI (Grad-CAM) - Open-access and research-friendly **Intended Use** - Tuberculosis detection using deep learning - Model benchmarking and comparison - Transfer learning and domain adaptation - Explainable AI research in medical imaging ⚠️ **Disclaimer:** *This dataset is intended for research and educational purposes only and should not be used for direct clinical diagnosis.*