Abstract

The development of robust artificial intelligence (AI) algorithms for breast imaging requires large, well-annotated datasets representative of diverse and underrepresented populations. However, publicly available mammography datasets predominantly originate from Western screening programs and provide limited representation of South Asian populations and screening-naive cohorts. We present the Indian Mammography DataBase (IMDB), an open-access mammography resource comprising two independently curated cohorts from a screening-naive Indian population undergoing diagnostic and opportunistic screening mammography. The release includes 12,674 mammography images from 3,219 patients across IMDB r1.0 and r2.0, making it one of the largest publicly available mammography datasets from India. Associated metadata include patient age, breast density, BI-RADS assessment, and histopathology-confirmed outcomes for suspicious lesions. Ground-truth labels were established using histopathology for BI-RADS 4/5 examinations and imaging / telephonic follow-up or triple expert readings for lower-risk examinations. Quality assurance procedures included image-view verification, metadata consistency checks, pathology linkage validation, and de-identification audits. As a representative AI use case, multiple convolutional neural network and YOLO-based models were successfully evaluated and trained using the dataset, demonstrating its suitability for machine-learning applications such as breast cancer classification, risk prediction, and algorithm benchmarking. The dataset is publicly available through the Indian Biological Images Archive (IBIA) and MIDAS, providing a FAIR-compliant resource for the development and evaluation of breast imaging AI systems.