Purification and Regulation: Comorbidity-Aware Multi-Label Few-Shot Learning for Medical Image Classification

Ying-Chih Lin; Po-Chih Kuo; Yong-Sheng Chen

cs.CVSep 18, 2026

Purification and Regulation: Comorbidity-Aware Multi-Label Few-Shot Learning for Medical Image Classification

Authors: Ying-Chih Lin, Po-Chih Kuo, Yong-Sheng Chen

Organizations: Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, 300093, Taiwan · Department of Computer Science, National Tsing Hua University, Hsinchu, 300093, Taiwan

Abstract

Multi-label few-shot learning (MLFSL) remains a significant challenge in medical image analysis (MIA). Current metric-based meta-learning methods face two critical limitations in MIA. First, conventional prototype generation often entangles irrelevant disease information, leading to contaminated prototypes and degraded performance. Second, prior studies typically enforce inter-class separability in embedding space, largely neglecting the inherent correlations among diseases. To overcome these challenges, we propose Prototype Purification and Regulation (PPR), a novel MLFSL framework for MIA. PPR first performs prototype purification by leveraging sample-level comorbidity scores to emphasize disease-specific features, producing purified prototypes that better characterize each disease. Building upon these purified prototypes, PPR further addresses the underexplored problem of inter-class prototype distance in MIA by incorporating disease-level comorbidity statistics to adaptively regulate inter-class similarity, forming a comorbidity-aware embedding space. Overall, PPR sequentially enables the model to capture pure disease features and inter-class relationships for reliable MLFSL in MIA. Extensive experiments across four chest X-ray benchmark datasets, including cross-domain evaluation, show that PPR consistently outperforms state-of-the-art methods, significantly improving disease detection while demonstrating robust generalization and clinical applicability.

Figures & tables

Figure 1: Comparison of prototype generation strategies in a 3-way 3-shot example. Gradient-color samples indicate mixed feature representations caused by multi-label co-occurrence, while single-colored samples correspond to pure single-disease features. (a) Support samples are encoded into the embedding space. (b) Existing methods derive a reference feature by simply averaging support features, then perform similarity-weighted averaging based on this biased reference, resulting in a contaminated prototype. (c) Our method enables the model to capture pure features by aligning the reference feature toward disease-specific representations, producing a purified prototype that better preserves disease characteristics in the embedding space.

Figure 2: Overview of the proposed PPR framework in a 3-way 3-shot example. (a) Purification: The reference feature is aligned toward purer representations by leveraging the label distribution derived from the sample comorbidity score as a purity prior, thereby emphasizing discriminative features and generating purified prototypes. (b) Regulation: prototype similarities are adaptively constrained based on disease comorbidity statistics to capture inter-disease relationships. (c) Classification: query features are concatenated with the purified prototypes and fed into the relation network for final prediction.

		5-shot
Dataset	Method	maF1	miF1	maR	miR	maP	miP	mAP
	ProtoNet+NLC [ 40 ]	34.75	37.13	36.73	36.73	38.75	38.73	39.54
	RelationNet [ 40 ]	41.24	44.11	60.22	60.23	36.03	36.02	46.46
	LPN+NLC [ 40 ]	32.17	37.06	35.60	35.60	37.52	37.54	39.45
	BCR [ 3 ]	46.12	46.78	70.38	70.38	35.51	35.51	46.62
CheXpert [ 17 ]	PPR (Ours)	49.82	50.41	86.84	86.85	35.72	35.72	46.37

Table 1: Results of overall performance comparison (%) in the 3-way 5-shot setting. The best scores are highlighted in bold and the second-best in blue.

		10-shot
Dataset	Method	maF1	miF1	maR	miR	maP	miP	mAP
	ProtoNet+NLC [ 40 ]	36.69	38.81	38.48	38.47	40.87	40.82	40.30
	RelationNet [ 40 ]	48.34	49.09	79.87	79.87	35.69	35.69	46.40
	LPN+NLC [ 40 ]	44.73	47.07	47.38	47.45	47.08	47.13	44.60
	BCR [ 3 ]	47.68	47.85	74.80	74.80	35.25	35.25	46.21
CheXpert [ 17 ]	PPR (Ours)	50.14	50.35	87.62	87.63	35.59	35.59	46.53

Table 2: Results of overall performance comparison (%) in the 3-way 10-shot setting. The best scores are highlighted in bold and the second-best in blue.

		CheXpert [ 17 ]			MIMIC [ 20 ]			CXR14 [ 45 ]
Method	shot	Pneumonia	Lung Lesion	Fracture	Enl. Card.	Lung Lesion	Fracture	Hernia	Edema	Pneumonia	Fibrosis
ProtoNet+NLC [ 40 ]		37.16	37.03	37.08	36.64	36.57	36.64	42.67	42.71	42.69	42.57
RelationNet [ 40 ]		45.01	45.11	45.10	33.70	33.60	33.76	35.95	35.99	35.87	35.95
LPN+NLC [ 40 ]		36.44	36.48	36.58	35.46	35.54	35.44	44.22	44.31	44.12	44.31
BCR [ 3 ]		47.16	47.21	47.21	44.04	44.09	44.02	36.20	36.57	35.92	36.57
PPR (Ours)	5	50.58	50.63	50.64	47.34	47.36	47.31	45.92	46.06	45.87	45.88

Table 3: Results of individual disease performance comparison using F1 score (%) in the 3-way 5-shot and 3-way 10-shot settings. The best scores are highlighted in bold and the second-best in blue.

Dataset	Method	maF1	miF1	maR	miR	maP	miP	mAP
	ProtoNet+NLC [ 40 ]	45.67	47.85	47.89	47.92	47.64	47.64	45.72
	RelationNet [ 40 ]	39.07	42.24	45.61	45.67	41.16	41.22	54.32
	LPN+NLC [ 40 ]	44.73	47.07	47.38	47.45	47.08	47.13	44.60
	BCR [ 3 ]	46.18	47.11	49.42	49.44	46.57	46.56	56.41
CheXpert [ 17 ] →CXR-LT [ 13 ]	PPR (Ours)	51.43	51.70	62.61	62.61	44.28	44.28	57.45
	ProtoNet+NLC [ 40 ]	40.19	37.23	34.87	36.92	36.91	36.81	39.78

Table 4: Performance comparison results of cross-domain rare disease classification (%) in the 3-way 5-shot setting. The best scores are highlighted in bold and the second-best in blue.

Dataset	Method	maF1	miF1	maR	miR	maP	miP	mAP
Chexpert [ 17 ]	Baseline (mean)	41.62	43.31	56.45	56.46	35.71	35.71	47.08
	Baseline (ws)	46.74	47.66	72.50	72.51	35.94	35.94	46.40
	Baseline (ws)+ $\mathcal{L}_{\text{pur}}$	48.27	48.98	78.26	78.28	35.77	35.78	46.93
	PPR	49.82	50.41	86.84	86.85	35.72	35.72	46.37
MIMIC [ 20 ]	Baseline (mean)	43.08	44.04	58.83	58.83	35.79	35.79	46.94
	Baseline (ws)	38.18	38.94	45.15	45.15	35.38	35.37	46.42

Table 5: Ablation study of loss functions and different prototype generation methods in the 3-way 5-shot setting (%).

Dataset	Method	maF1	miF1	maR	miR	maP	miP	mAP
Chexpert [ 17 ]	max	47.09	48.47	76.13	76.14	35.83	35.83	46.55
	min	47.25	48.24	75.26	75.27	35.78	35.78	46.69
	PPR	49.82	50.41	86.84	86.85	35.72	35.72	46.37
MIMIC [ 20 ]	max	40.51	42.6	55.53	55.53	35.52	35.51	46.82
	min	42.54	43.66	58.54	58.54	35.74	35.73	46.49
	PPR	46.27	47.08	70.87	70.88	35.54	35.55	46.67

Table 6: Ablation study of different prototype regulation methods in the 3-way 5-shot setting (%).

Figure 3: Visualization of representations trained on CheXpert in the 3-way 5-shot setting. The radius indicates intra-class compactness, with bars closer to the circumference denoting higher intra-class feature similarity. The angle reflects inter-class distance, where a wider angle corresponds to a greater distance between disease classes. The dotted line denotes the comorbidity statistics $j$ between two diseases used in training. Base and novel classes are displayed in the upper and lower plots, respectively.

Figure 4: Hyperparameter analysis of

\delta

on CheXpert in the 3-way 5-shot setting.

Figure 5: Hyperparameter analysis of

\lambda

on CheXpert in the 3-way 5-shot setting.

Few-Shot Learning

Few-shot learning develops methods that recognize or adapt to new tasks from only a handful of labeled examples.

See all 213 papers

One Round Is All You Need: Analytic Federated Learning for Task-Heterogeneous Multi-Label Medical Image Classification
Jul 22, 2026Afsaneh Mahanipour, Hana KhamfroushChest X-Ray Classification Federated Learning Aggregation
Department of Computer Science University of Kentucky Lexington, KY, USA
Early Intervention for VFM-based Multimodal Medical Image Classification
Mar 18, 2026Qijie Wei, Hailan Lin, Xirong LiCross-Modal Representation Learning Vision Foundation Model Adaptation
Renmin University of China, Beijing 100872, China
PRIMA: Pre-Training with Risk-Integrated Image--Metadata Alignment for Medical Diagnosis with LLM-Based Feature Aggregation
Feb 26, 2026Yiqing Wang, Chunming He, Ziyun Yang +5Cross-Modal Alignment Multimodal Pretraining
Department of Biomedical Engineering, Duke University 101 Science Drive, Durham, NC 27705 USA · Department of Ophthalmology and Visual Sciences, University of Michigan 1000 Wall Street, Ann Arbor, MI, USA

Jul 22, 2026J/K move · Enter open · S save

One Round Is All You Need: Analytic Federated Learning for Task-Heterogeneous Multi-Label Medical Image Classification

Afsaneh Mahanipour, Hana Khamfroush

Department of Computer Science University of Kentucky Lexington, KY, USA

Chest X-Ray Classification Federated Learning Aggregation Medical Image Classification Federated Learning Heterogeneous Federated Learning

Open paper AskPDF arXiv

Federated learning (FL) enables multiple clinical institutions to collaboratively train a shared disease classifier without centralizing patient data. In practice, however, each institution annotates only the pathologies within its area of expertise, so the federation operates under task heterogeneity: each client holds labels for a strict subset of the target disease categories while the remaining classes are entirely unobserved at that site. Existing gradient-based FL methods fail under this setting because they require hundreds of communication rounds to converge and because missing class labels introduce systematic false-negative bias that the model cannot correct without a principled mechanism. We propose an analytic federated learning framework for multi-label medical image classification under task heterogeneity. The proposed method replaces iterative gradient optimization with three closed-form operations: a balanced label projection that neutralizes class-imbalance bias by normalizing positive and negative contributions to equal total mass; a per-class absolute aggregation law that independently assembles the optimal ridge-regression classifier for each disease category from the sufficient statistics uploaded by its annotating clients; and an optional analytic pseudo-label refinement round that propagates missing-class knowledge from a confidence-filtered teacher classifier to non-annotating clients. The entire procedure requires at most two communication rounds, irrespective of the degree of task heterogeneity or the number of participating clients. Experiments on ChestXray14 under four progressively severe missing-class configurations demonstrate that the proposed method consistently outperforms the state-of-the-art federated multi-label method FedMLP by up to 18.44 BACC points and 13.24 AUC points, while reducing the communication.

Abstract

Few-Shot Learning

Explore similar work

One Round Is All You Need: Analytic Federated Learning for Task-Heterogeneous Multi-Label Medical Image Classification

Early Intervention for VFM-based Multimodal Medical Image Classification

PRIMA: Pre-Training with Risk-Integrated Image--Metadata Alignment for Medical Diagnosis with LLM-Based Feature Aggregation