Cross-domain self-supervised training towards universal representation learning in medical imaging

In recent years, the medical field has benefited greatly from the use of machine learning technologies and Deep Learning in particular. More and more tasks are now being performed using these techniques, such as classification and segmentation of lesions, in order to drastically reduce the radiologist's time and effort. Although these algorithms are extremely effective, they require a huge amount of data to generalize due to the diversity of modalities, organs, acquisition devices and clinical tasks. This problem is not trivial, considering that the data must be tagged, which requires a huge investment of time and resources by doctors or radiologists. A common approach to solving this problem is transfer learning. This is a technique that uses models that have been previously trained on ImageNet, a dataset of approximately 1 million naturalistic images. Although this solution is widely used in the medical field, there are several studies that question whether this choice could lead to suboptimal results due to the extreme gap between these domains. In this work, alternative solutions are explored, specifically the use of self-supervised learning techniques. Self-supervised learning is the process by which a model is able to learn intrinsic properties of the data without necessarily being guided by previously assigned labels. For the first phase of this work, several public datasets were selected and then combined into a single dataset. The goal of this process was to create a dataset comparable in size to ImageNet, but consisting of images collected in the medical domain. The training slices were sampled from the SSL dataset to make it balanced, and then preprocessed per modality to obtain a dataset with a uniform scale of values and image size. Self-supervised training was performed using a recent contrastive technique, DINO, in which different augmented views of the same image are compared to produce similar features for all input data. Therefore, it was important to investigate which transformations are most appropriate in this domain. Finally, two downstream tasks were selected to verify the effectiveness of the self-supervised training and evaluate its performance. The results obtained show that this type of training can be used to train deep learning models that can be effectively transferred to clinical tasks involving different body parts and modalities.