Semi-Supervised Learning with Pseudo-Labeling for Pancreatic Cancer Detection on CT Scans

Deep learning techniques have recently gained increasing attention not only among computer science researchers but are also being applied in a wide range of fields. However, deep learning models demand huge amounts of data. Furthermore, fully supervised learning requires labeled data to solve classification, recognition, and segmentation problems. Data labeling and annotation in the medical domain are time-consuming and labor-intensive. Semi-supervised learning has demonstrated the ability to improve deep learning performance when labeled data is scarce. However, it is still an open and challenging question on how to leverage not only labeled data but also the huge amount of unlabeled data. In this paper, the problem of pancreatic cancer detection on CT scans is addressed by a semi-supervised learning approach based on pseudo-labeling. Preliminary results are promising and show the potential of semi-supervised deep learning to detect pancreatic cancer at an early stage with a limited amount of labeled data.