Ranking-Guided Semi-Supervised Domain Adaptation for Severity Classification

1. Problem Statement

The paper addresses the challenge of Domain Shift in medical image analysis, specifically for severity classification (ordinal classification).

• Context: In medical settings (e.g., ulcerative colitis or diabetic retinopathy), classifiers trained on labeled data from a source hospital often fail when applied to a target hospital due to differences in imaging equipment, protocols, and conditions.
• Specific Challenge: Unlike standard classification where classes are distinct clusters, severity classification involves ordered, continuous class labels (e.g., mild, moderate, severe). The boundaries between these classes are inherently ambiguous.
• Limitation of Existing Methods: Traditional Semi-Supervised Domain Adaptation (SSDA) methods assume distinct class clusters in feature space. When applied to severity classification, these methods fail because they cannot capture the ordinal nature of the data, leading to misaligned feature distributions and suboptimal performance.

2. Methodology

The authors propose a novel framework that aligns source and target domains by learning a continuous rank score rather than treating classes as discrete, independent categories. The network architecture consists of a feature extractor ($F_e$) branching into a classifier ($F_c$) and a ranking function ($F_r$).

The method is optimized in two main stages using two core components:

A. Cross-Domain Ranking (CDR)

• Goal: To establish a consistent severity scale across domains by learning relative rankings between source and target samples.
• Mechanism: Instead of ranking only within a single domain, CDR forms pairs of labeled samples from both the source ($D_s$) and target ($D_t$) domains.
• Loss Function: It minimizes a pairwise ranking loss where the relative order of the pair is determined by their ground-truth class labels ($y_i > y_j$).
• Effect: This forces the model to learn a ranking function $r(x)$ that produces consistent rank scores for samples of the same relative severity, regardless of which domain they originate from. This aligns the global distribution of severity levels across domains.

B. Continuous Distribution Alignment (CDA)

• Goal: To align the rank score distributions of the source and target domains, specifically handling unlabeled target samples ($D_u$).
• Mechanism:
  1. Soft Labeling: For unlabeled target samples, the method estimates the probability of belonging to each class using a Gaussian Mixture Model (GMM) based on the current rank score distribution. This creates rank-based soft labels ($w_{i,k}$) rather than hard pseudo-labels.
  2. Alignment: The method minimizes the distance between the rank score of a sample and the mean rank score (prototype $\mu_k$) of the corresponding class in the source domain.
• Effect: By treating classes as continuous distributions rather than discrete points, CDA mitigates the issue of ambiguous class boundaries. It ensures that the target domain's feature distribution aligns with the source domain's ordinal structure.

Overall Objective

The total loss function combines:

1. Classification loss ($L_c$) on labeled data.
2. Pairwise ranking loss ($L_r$) on labeled data (both intra-domain and cross-domain).
3. Continuous Distribution Alignment loss ($L_a$) on all data (labeled and unlabeled).

3. Key Contributions

1. Ranking-Guided SSDA Framework: A novel approach for severity classification that leverages learning-to-rank to handle continuous class boundaries, moving away from the assumption of distinct class clusters.
2. Cross-Domain Ranking (CDR): A mechanism that learns rankings from sample pairs across different domains to ensure a consistent severity interpretation and align rank score distributions.
3. Continuous Distribution Alignment (CDA): A loss function utilizing rank-based soft labels to align the continuous severity distributions of unlabeled target samples with the source domain, effectively handling class ambiguity.

4. Experimental Results

The method was evaluated on two medical imaging tasks:

• Ulcerative Colitis (UC): Endoscopic images (LIMUC dataset $\to$ Private Hospital dataset).
• Diabetic Retinopathy (DR): Fundus images (DDR dataset $\to$ FGADR dataset).
• Setup: Semi-supervised setting with only 10 labeled target samples per class.

Performance Highlights:

• Superiority: The proposed method outperformed state-of-the-art baselines (including MME, CDAC, SLA, and ORUDA) in terms of Mean Macro Precision (mP) and Mean Macro F1 (mF1).
• Ablation Study:
  • Adding CDR significantly improved Precision (better class discrimination).
  • Adding CDA significantly improved Recall (better coverage of the distribution).
  • Using both components together yielded the best overall performance, demonstrating their complementary nature.
• Visualization: t-SNE and rank distribution plots confirmed that the method successfully aligned the rank score distributions between source and target domains for each severity class, whereas baseline methods showed misalignment.

5. Significance

• Medical Relevance: This work provides a robust solution for deploying diagnostic AI in new hospitals where labeled data is scarce. It specifically addresses the nuance of disease progression, which is often continuous rather than categorical.
• Theoretical Advancement: It challenges the standard domain adaptation assumption of "distinct clusters," proposing instead that ordinal structure should be preserved and aligned via rank scores.
• Practical Impact: By effectively utilizing a small number of labeled target samples and many unlabeled ones, the method reduces the annotation burden for medical experts while maintaining high diagnostic accuracy across different imaging environments.