Self-Supervision for Medical Image Classification: State-of-the-Art Performance with ~100 Labeled Training Samples per Class.

Interestingly and despite existing large annotated data sets like ImageNet (currently more than 14 million images, [ 5 ]) or the COCO data set (more than 200,000 labeled images, [ 6 ]) and impressive performance of supervised DL in corresponding benchmarks, recent developments in the natural image and computer vision domain showed a trend toward self-supervised learning (SSL).

also in 53,322 other papers