An open question for certain tasks in the Medical AI and Computer Vision space is whether pre-training on domain-specific data yields a statistically significant improve- ment in performance as opposed to pre-training on non-domain-specific data. One such task where this question is yet unanswered is that of diagnosing COVID19 from chest Computed Tomography (CT) scans. A further question is what pre-training methods might be effective in bringing about this improvement. In this paper, we show that for the downstream task of COVID19 classification, pre-training a ResNet50 model with the self-supervised learning technique of DINO on an unlabelled domain-specific data set of fewer than 100 000 CT images yields a statistically significant improvement over using a model with pre-trained ImageNet weights. The Area Under the Receiver Op- erating Characteristic (ROC) Curve (AUC) is used to quantify this performance, while Permutation Testing is used to show that the increase is statistically significant.

We use slurm to run our experiments, see this sbatch file for reference.