MVCNet: Multiview Contrastive Network for Unsupervised Representation Learning for 3-D CT Lesions

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for computed tomography (CT) diagnostics requires extensive lesion-level annotations, which are costly and scarce.
• Existing methods may underutilize the 3-D spatial information present in CT data.

Purpose of the Study:

• To develop a novel multiview contrastive network (MVCNet) for enhancing CT data representation learning.
• To improve the utilization of scarce annotated CT data by leveraging 3-D information more effectively.

Main Methods:

• MVCNet processes each 3-D lesion from multiple 2-D orientations, learning contrastive representations.
• A contrastive loss function aggregates views of the same lesion and separates views of different lesions.
• Uninformative negative samples are filtered to enhance feature discriminability.

Main Results:

• MVCNet achieved state-of-the-art unsupervised representation learning accuracies on LIDC-IDRI (88.62%), LNDb (76.69%), and TianChi (84.33%) datasets.
• When fine-tuned on only 10% of labeled data, MVCNet performance was comparable to fully supervised models.
• This demonstrates MVCNet's superiority in learning from limited annotations.

Conclusions:

• Contrasting multiple 2-D views effectively captures 3-D information from CT scans.
• MVCNet significantly improves the utility of scarce annotated CT data for diagnostic algorithms.