Integrations: Informed Subset Selection Based Generation for Medical Imaging in Resource Constrained Setting

Deep learning models are revolutionizing the field of radiology and diagnostic medicine. However, these models require a large amount of labelled data causing an impediment to their utilization in low resource setting. In this paper (INTEGRATIONS, INformed subseT sElection based GeneRATion for medical Imaging in resource cONstrained Setting), we demonstrate that informed subset selection followed by human labelling and then Wasserstein GANs (wGANs) driven data augmentation performed comparably to large labelled data-based prediction. This demonstrates that despite having less amount of labelled data, one can reach the required quantum of large labelled data by data augmentation using wGANs, albeit if a small subset (10%) of images selected for human labelling is performed using informed subset selection.