LVCL: Label-Volume Contrastive Learning for Multi-Label Classification of Retinal Oct Volumes

Optical coherence tomography (OCT) technique can produce volumetric data of the retina for disease diagnosis. Each OCT volume consists of 2D B-scans, and recent studies have shown remarkable success with deep learning for single-label B-scan classification tasks. However, B-scan annotation is quite difficult and single-label classification approaches cannot meet the growing clinical demands. It is attractive to develop a multi-label classification approach for disease diagnosis only using volume-level labels. In this paper, we propose a label-volume contrastive learning (LVCL) for the multi-label classification of OCT volumes. In LVCL, a robust volume encoder (RVE) is proposed to convert an OCT volume into its volume embedding by modeling the local and global dependencies among B-scans. Meanwhile, a label encoder outputs discriminative label embeddings for all pathological labels. Lastly, the correspondence between volume embedding and label embeddings is learned by contrastive learning and our proposed loss function. Experiments on 5121 OCT volumes of 1244 patients show superior multi-label volume classification performance than other approaches.