Automatic grading of cervical biopsies by combining full and self-supervision

In computational pathology, the application of Deep Learning to the analysis of Whole Slide Images (WSI) has provided results of unprecedented quality. Due to their enormous size, WSIs have to be split into small images (tiles) which are first encoded and whose representations are then agglomerated in order to solve prediction tasks, such as prognosis or treatment response. The choice of the encoding strategy plays a key role in such algorithms. Current approaches include the use of encodings trained on unrelated data sources, full supervision or self-supervision. In particular, self-supervised learning (SSL) offers a great opportunity to exploit all the unlabelled data available. However, it often requires large computational resources and can be challenging to train. On the other end of the spectrum, fully-supervised methods make use of valuable prior knowledge about the data but involve a costly amount of expert time. This paper proposes a framework to reconcile SSL and full supervision and measures the trade-off between long SSL training and annotation effort, showing that a combination of both has the potential to substantially increase performance. On a recently organized challenge on grading Cervical Biopsies, we show that our mixed supervision scheme reaches high performance (weighted accuracy (WA): 0.945), outperforming both SSL (WA: 0.927) and transfer learning from ImageNet (WA: 0.877). We further provide insights and guidelines to train a clinically impactful classifier with a limited expert and/or computational workload budget. We expect that the combination of full and self-supervision is an interesting strategy for many tasks in computational pathology and will be widely adopted by the field.