Multi-Label Radiology AI: Why Dominant-Label Supervision Fails in Singapore Hospitals

Multi-Label Radiology AI: Why Dominant-Label Supervision Fails in Singapore Hospitals

Most radiology AI models deployed in Singapore hospitals are trained on datasets where each image receives a single diagnostic label—the most severe finding. A chest X-ray showing both pneumonia and pleural effusion gets labeled "pneumonia." A fundus image with diabetic retinopathy and glaucoma gets tagged "severe diabetic retinopathy." This dominant-label supervision is efficient for annotation but catastrophic for clinical deployment: the model learns to suppress evidence of coexisting conditions, exactly when clinicians need comprehensive differential diagnosis support.

This post is for hospital AI leads, clinical informatics teams, and radiology department heads evaluating computer vision AI for deployment in Singapore health systems. We examine why standard weak supervision fails for multi-pathology imaging, what recent research reveals about dominant-label multiple instance learning, and how to structure procurement and validation to avoid silent failures in production.

Key takeaways

• Dominant-label supervision suppresses coexisting lesions: Standard multiclass training treats the most severe diagnosis as a complete one-hot description, penalizing models that detect lower-severity coexisting conditions [3]
• Multiple instance learning (MIL) recovers hidden evidence: Formulating radiology AI as dominant-label MIL allows models to identify all lesions present, not just the annotated dominant label [3]
• Multi-label architectures reduce false negatives: Lightweight hybrid models designed for imbalanced fundus datasets achieve better sensitivity across rare coexisting conditions than single-label classifiers [4]
• RAG-augmented vision-language models improve factual consistency: Gated multimodal retrieval can reduce hallucinations in radiology report generation, but retrieval quality depends on joint image-text semantics [2]
• Singapore deployment requires explicit multi-label validation: Procurement specs must test model performance on images with multiple confirmed pathologies, not just dominant-label test sets

Why dominant-label supervision creates silent failures

In Singapore public hospitals, radiology AI procurement typically specifies AUROC targets on held-out test sets. Vendors train on large datasets—often international cohorts where each image is labeled with the most clinically significant finding. A recent preprint on whole-slide pathology imaging formalizes this as dominant-label multiple instance learning: the observed label is the maximum-severity element of a latent lesion set [3].

Standard multiclass MIL treats this dominant label as ground truth and penalizes the model whenever it activates on lower-severity regions. The result: models learn to ignore early-stage disease, incidental findings, and coexisting pathologies. In production, this manifests as:

• Missed secondary diagnoses: A pneumonia-detection model that suppresses pleural effusion evidence because training images with both conditions were labeled "pneumonia"
• False confidence in single-pathology reports: The model outputs high confidence for the dominant label and low confidence for everything else, even when multiple conditions are present
• Audit trail gaps: When a radiologist later identifies a coexisting lesion the AI missed, the model's internal attention maps show it saw the region but was trained to suppress it

We've seen this failure mode in vendor pilots at Singapore hospital clusters: models with excellent single-label AUROC that systematically miss secondary findings clinicians expect AI to flag.

How multiple instance learning recovers coexisting lesions

The dominant-label MIL framework reframes the problem: instead of treating the observed label as complete, it models the label as the maximum of a latent set of lesion severities [3]. During training, the model learns to:

1. Identify all candidate lesion regions (instances) in the image (bag)
2. Predict severity for each region independently
3. Aggregate predictions such that the bag-level output matches the dominant label, but instance-level activations are preserved

This approach allows the model to detect coexisting lower-severity lesions without contradicting the training label. In pathology whole-slide imaging, this recovers evidence of multiple tumor grades and precancerous lesions that standard MIL suppresses [3].

For radiology deployment in Singapore hospitals, this means:

• Procurement specs should require multi-instance architectures for use cases where coexisting pathologies are common (chest X-ray, fundus imaging, abdominal CT)
• Validation datasets must include confirmed multi-pathology cases, not just dominant-label test sets sampled from the same distribution as training data
• Model cards should disclose training label methodology: Was each image labeled with all findings, or only the most severe?

Multi-label architectures for imbalanced radiology datasets

Even when training labels include multiple diagnoses, class imbalance creates deployment risk. A recent preprint on fundus imaging introduces LightHybridNet, a lightweight hybrid architecture designed for multi-label classification in imbalanced datasets [4]. Key design choices:

• Multi-label loss functions that don't assume mutual exclusivity between diagnoses
• Class-balanced sampling during training to prevent rare conditions from being ignored
• Lightweight architecture (suitable for edge deployment in Singapore polyclinics with limited GPU infrastructure)

For Singapore hospitals evaluating radiology AI, this highlights a procurement gap: most vendor RFPs specify single-label accuracy metrics (sensitivity, specificity, AUROC for each condition independently). Multi-label performance requires:

• Subset accuracy: Percentage of images where all present labels are correctly predicted
• Hamming loss: Average per-label error rate across the label set
• Example-based F1: Harmonic mean of precision and recall computed per image, then averaged

Without these metrics, vendors can game single-label benchmarks while delivering poor multi-pathology performance in production.

RAG-augmented vision-language models for radiology reports

Beyond lesion detection, Singapore hospitals are piloting vision-language models for radiology report generation and clinical question answering. A recent preprint on medical visual question answering (MVQA) shows that retrieval-augmented generation (RAG) reduces factual inconsistencies in Med-LVLMs, but only when retrieval is guided by joint image-text semantics [2].

The paper introduces gated multimodal representations: instead of retrieving evidence based on the text query alone, the system encodes both the image and question, retrieves relevant clinical cases, and gates the retrieved context based on multimodal relevance. This reduces hallucinations in radiology QA tasks [2].

For Singapore deployment, this matters because:

• Text-only RAG misses visually similar cases: A query about "ground-glass opacities" retrieves text descriptions but may miss radiologically similar images with different terminology
• Multimodal retrieval requires curated image-report databases: Singapore hospitals need structured archives of de-identified cases with high-quality annotations, not just PACS dumps
• Gating mechanisms prevent irrelevant retrieval from degrading answers: Retrieved cases must be filtered for joint visual and semantic relevance, not just keyword overlap

We've worked with Singapore health systems to structure RAG evaluation for radiology LLMs; see our clinical AI services for governance frameworks and validation protocols.

Why this matters in Singapore

Singapore's radiology AI adoption is accelerating: multiple hospital clusters have deployed chest X-ray AI, and fundus screening AI is expanding in polyclinics under the National Diabetic Retinopathy Screening Programme. But procurement processes inherited from traditional software acquisition don't capture multi-label failure modes:

• Vendor demos use single-pathology test cases that don't stress coexisting-lesion performance
• Clinical validation studies report per-condition sensitivity/specificity without testing multi-pathology images
• Post-deployment monitoring tracks overall accuracy but doesn't flag systematic suppression of secondary findings

The result: models that pass procurement validation but create clinical risk in production. A pneumonia-detection AI that misses pleural effusion doesn't fail catastrophically—it just shifts cognitive load back to radiologists, exactly where AI was supposed to help.

Singapore's HSA AI-SaMD exemption pathway for public healthcare deployment requires "clinical evidence of safety and performance," but current guidance doesn't specify multi-label validation requirements. Hospital AI governance teams need to fill this gap in procurement specs and post-market surveillance.

What to do next

If you're evaluating radiology AI for deployment in Singapore hospitals:

1. Audit vendor training data for dominant-label bias: Request disclosure of labeling methodology—were images annotated with all findings, or only the most severe? If dominant-label, require evidence that the model architecture recovers coexisting lesions.
2. Add multi-label metrics to procurement specs: Require subset accuracy, Hamming loss, and example-based F1 on a validation set with confirmed multi-pathology cases from your institution.
3. Test on local multi-pathology cases before go-live: Don't rely on vendor-provided test sets. Curate 50–100 cases from your PACS with radiologist-confirmed coexisting findings and measure per-lesion sensitivity.
4. Structure post-deployment monitoring for secondary-finding misses: Track not just overall accuracy but specifically cases where the model flagged the dominant pathology but missed a coexisting condition later identified by radiologists.
5. For RAG-augmented report generation, validate retrieval quality: If deploying vision-language models for radiology QA, test whether retrieved cases are visually and semantically relevant, not just keyword-matched. See our open-source clinical analytics governance frameworks for RAG evaluation protocols.

Need help structuring multi-label validation for radiology AI procurement? Start a project with our clinical AI deployment team.

FAQ

What is dominant-label multiple instance learning?

Dominant-label MIL is a weakly supervised learning framework where each training image (bag) is labeled with only the most severe diagnosis, even though multiple lesions (instances) may be present. The model learns to identify all lesions by treating the observed label as the maximum of a latent lesion set, rather than a complete description [3].

Why do standard radiology AI models suppress coexisting lesions?

Standard multiclass training treats the dominant label as ground truth and penalizes the model for activating on other regions. This teaches the model to ignore lower-severity findings, even when they're clinically significant. The model learns "if pneumonia is present, suppress pleural effusion evidence" because training images with both were labeled only "pneumonia" [3].

How should Singapore hospitals validate multi-label radiology AI?

Procurement validation should include: (1) a test set of local cases with radiologist-confirmed multiple pathologies, (2) multi-label metrics (subset accuracy, Hamming loss, example-based F1), and (3) per-lesion sensitivity analysis to detect systematic suppression of secondary findings. Single-label AUROC is insufficient.

Can RAG improve radiology report generation accuracy?

Yes, but only with multimodal retrieval. Text-only RAG retrieves cases based on keywords but misses visually similar images. Gated multimodal retrieval—encoding both the image and query, then filtering retrieved cases for joint relevance—reduces hallucinations in radiology QA tasks [2]. Singapore hospitals need curated image-report databases for effective RAG deployment.

Sources

[1] Assessment of change in drug-induced parkinsonism in daily-life gait using vision-based machine learning. PLOS Digital Health, October 15, 2026. https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001780

[2] Enhancing RAG-based Medical VQA with Gated Multimodal Representations. arXiv preprint, October 16, 2026. https://arxiv.org/abs/2610.23588v1

[3] Beyond the Dominant Label: Recovering Coexisting Lesion Evidence with Multiple Instance Learning. arXiv preprint, October 15, 2026. https://arxiv.org/abs/2610.22091v1

[4] LightHybridNet: A Lightweight Hybrid Deep Learning Framework for Efficient Multi-Label Ophthalmic Disease Classification in Imbalanced Fundus Images. arXiv preprint, October 14, 2026. https://arxiv.org/abs/2610.20671v1