Deep Learning Approaches for Medical Imaging Under Varying Degrees of Label Availability: A Comprehensive Survey

TL;DR

This survey addresses the challenge of medical imaging with varying label availability by organizing deep learning approaches into incomplete, inexact, and absent supervision. It synthesizes around 600 studies (2018–2024) across classification, segmentation, and detection to present a unified view of active learning, semi-supervised learning, MIL, unsupervised learning, and transductive transfer learning. Key contributions include formal definitions, taxonomies, and cross-paradigm integrations (e.g., AL with Semi-SL, UL with TTL), plus practical insights into datasets, modalities, and tasks. The work highlights future directions such as multi-modal integration and foundation-model pretraining to further reduce labeling costs while improving robustness and generalization in clinical settings.

Abstract

Deep learning has achieved significant breakthroughs in medical imaging, but these advancements are often dependent on large, well-annotated datasets. However, obtaining such datasets poses a significant challenge, as it requires time-consuming and labor-intensive annotations from medical experts. Consequently, there is growing interest in learning paradigms such as incomplete, inexact, and absent supervision, which are designed to operate under limited, inexact, or missing labels. This survey categorizes and reviews the evolving research in these areas, analyzing around 600 notable contributions since 2018. It covers tasks such as image classification, segmentation, and detection across various medical application areas, including but not limited to brain, chest, and cardiac imaging. We attempt to establish the relationships among existing research studies in related areas. We provide formal definitions of different learning paradigms and offer a comprehensive summary and interpretation of various learning mechanisms and strategies, aiding readers in better understanding the current research landscape and ideas. We also discuss potential future research challenges.

Figures (10)

• Figure 1: Mindmap of Learning Patterns Under Varying Degrees of Label Availability.
• Figure 2: Here we use the Cardiac MRI segmentation task as an example to show different learning scenarios. In supervised learning, fully labeled data is given. In active learning (Section \ref{['Active Learning (AL)']}), a small amount of labeled data is combined with a large amount of unlabeled data and human annotation is used as additional assistance. Semi-supervised learning (Section \ref{['Semi-supervised Learning (Semi-SL)']}) is a similar scenario. In inexact supervised learning (Section \ref{['sec: Inexact Supervision']}), only scribble labels are available. Unsupervised learning (Section \ref{['Unsupervised Learning']}) only involves unlabeled data. Transductive transfer learning (Section \ref{['Transductive Transfer Learning']}) focuses on the labeled data from a different domain, and we use Cardiac MRI from another domain as an example here.
• Figure 3: The taxonomy of Active Learning.
• Figure 4: The taxonomy of Semi-supervised Learning.
• Figure 5: The taxonomy of Inexact-supervised Learning.

Theorems & Definitions (10)

• Definition 2.1: Classification
• Definition 2.2: Segmentation
• Definition 2.3: Object Detection
• Definition 2.4: Incomplete Supervision
• Definition 2.5: Inexact Supervision
• Definition 2.6: Absent Supervision
• Definition 3.1: Active Learning
• Definition 3.2: Semi-supervised Learning
• Definition 5.1: Unsupervised Learning
• Definition 5.2: Transductive Transfer Learning