Contrastive learning for annotation enrichment and automatic classification of pulmonary embolisms

Pulmonary embolism (PE) is a disease that can lead to death if not treated quickly. Although early diagnosis is critical, it is also challenging and sometimes missed. Therefore, deep learning methods are starting to be used for automatic PE detection. However, most common methods are limited by their need for a large and accurately labeled dataset. This master thesis develops two main axes to overcome this limitation. The first one aims to complete a partially annotated dataset, while the second one studies a novel method – contrastive learning - that achieves promising results with fewer labeled data. Both axes are developed in an annotation enrichment task, and this last one is implemented with both supervised and contrastive learning. This task involves training a network with fewer and fewer data to estimate the minimum number of samples needed to annotate the others with accuracy. The results, although still theoretical, are encouraging. With supervised learning, the annotation work is almost reduced by half. Only 52.97% of the dataset needs to be labeled or verified while the network completes the annotation of the other half with perfect accuracy (reached considering the labels of this remaining part of the dataset). Contrastive learning allows improving this result by achieving a percentage of required labeled data of 49.34%. This thesis shows that contrastive learning can outperform supervised learning and leads the way to further experiments more focused on PE detection purposes.