Learning with Limited Labels For Medical Image Analysis

Despite the tremendous progress, the need to have large-scale annotated data to train neural network models is, at best, undesirable, and improving the representation power of these models for medical images while preserving intrinsic properties of volumetric medical image modalities is yet under-explored. Obtaining well-annotated labels for medical images requires professional experts and heedful manual labeling, which is expensive and laborious, especially for 3D medical modalities (e.g., MRI and CT). This research addresses the necessity of designing models capable of learning intrinsic properties with limited labels, which is surprisingly not well-explored in the literature.