Beyond Multi-Stage Pipelines: A Label-Correlation Ensemble for Multi-Label Retinal Fundus Disease Classification

Md Asif Uddin; Jannatul Ferdous; Asif Islam; Zobaer Mahmud Zisan; Promit Ghosh; Shah Ahafik Arman; Niloy Farhan; Rafeed Rahman

cs.CVOct 16, 2026

Beyond Multi-Stage Pipelines: A Label-Correlation Ensemble for Multi-Label Retinal Fundus Disease Classification

Authors: Md Asif Uddin, Jannatul Ferdous, Asif Islam, Zobaer Mahmud Zisan, Promit Ghosh, Shah Ahafik Arman, Niloy Farhan, Rafeed Rahman

Organizations: Department of Computer Science and Engineering, BRAC University, Dhaka, Bangladesh

Abstract

Retinal fundus images frequently carry more than one disease, and public fundus datasets are heterogeneous and severely long-tailed. Clinical triage suggests a staged design, in which a screening gate, a disease-group router and group-specific classifiers operate in sequence, but it is unclear whether this hierarchy earns its complexity. We present a controlled comparison of three architectures on a unified 39-target label space built from RFMiD 1.0, RFMiD 2.0 and ODIR-5K (12,652 images): (i) a three-stage gated cascade of eight models, (ii) a flat model with 39 independent sigmoid outputs, and (iii) a label-correlation model that replaces the flat linear head with 39 learned label queries and a two-layer transformer decoder. All three share one Swin-V2-B backbone, one preprocessing pipeline and one training protocol, so differences are attributable to the architecture. Staging offers no consistent advantage: tracing every missed disease instance to the stage that lost it shows that the two coarse gates discard 34.2% of true instances while the fine-grained classifiers lose only 6.7%, and the gates' hard zeros depress the cascade's macro-AUC to 0.747. Label-correlation modeling gives the largest single gain, raising the detection rate from 0.621 to 0.682 (0.689 with test-time augmentation). Blending the flat and label-correlation models under four-view test-time augmentation is the best configuration on six of eight metrics (exact match 0.694, micro-F1 0.678, macro-F1 0.456, macro-AUC 0.972). On definition-matched class subsets it matches or exceeds four recent studies on macro- and micro-F1, most clearly on low-support pathologies. We also show that ensembling a gated cascade with continuous-output models is harmful, and that per-source negative sampling is needed to remove a dataset shortcut that otherwise yields a spurious AUC of 1.000.

Figures & tables

Source	Train	Val	Test	Native labels
ODIR-5K (per eye) [ 7 ]	6,092	861	1,642	8 categories + keywords
RFMiD 1.0 [ 4 ]	1,920	640	640	46 conditions
RFMiD 2.0 [ 6 ]	508	177	172	49 conditions
Merged	8,520	1,678	2,454	39 scored targets
of which healthy	3,652	617	1,005

TABLE I: Merged dataset. Each source keeps its official split.

			Stage 2		Stage 3
Group	Diseases	Train	AUC	AP	AUC	AP
G1 Vascular	DR HR HTN BRVO CRVO TV CWS PRH MCA	2,554	.970	.969	.971	.673
G2 Exudative	ARMD DN EDN CSR CME RPEC MHL ERM PT	698	.932	.775	.926	.582
G3 Optic disc	ODC GLA ODP ODE AION TD ST	806	.902	.745	.942	.660
G4 Structural	MYA TSLN LS RT RD CF	694	.978	.933	.983	.919
G5 Media	MH CAT	656	.969	.874	.983 †	.948 †

TABLE II: Disease groups, training images per group and validation performance of the Stage-2 router (per group) and the Stage-3 classifiers (macro over the group).

Fig. 1: The four systems compared. (A) Healthy images exit at Stage 1; unhealthy images are routed to every group that fires, and each routed group’s classifier makes the fine-grained call. (B) One linear head predicts all 39 targets. (C) 39 label queries attend to the image tokens and to each other. (D) The best configuration blends (B) and (C) under four-view TTA.

Model	Out.	Images	LR	Batch	Ep.	Pat.	Monitor
Stage 1	2 s	8,520	$10^{-5}$	8	60	8	AUC
Stage 2	7	4,868	$2{\cdot}10^{-5}$	8	30	7	macro-AUC
Stage 3 ( $\times$ 6)	2–9	162–2,554	$2{\cdot}10^{-5}$	8	30	7	macro-AP
Flat	39	8,520	$2{\cdot}10^{-5}$	8	30	7	macro-AP
Label-corr.	39	8,520	$2{\cdot}10^{-5}$	16	25	6	macro-AP

TABLE III: Training settings. All models: Swin-V2-B at 384 px, AdamW, weight decay

10^{-2}

, warm-up then cosine decay, batch size 4 with gradient accumulation (Batch: effective batch). s Softmax; all other heads are sigmoid.

Threshold $\tau_{1}$	.10	.20	.30	.50	.70	.90
Unhealthy recall	.896	.876	.862	.845	.825	.770
Healthy recall	.799	.853	.887	.912	.940	.968
Accuracy	.861	.867	.871	.870	.867	.843

TABLE IV: Stage-1 operating point on validation (1,061 unhealthy, 617 healthy images).

Stage 1 accuracy / AUC	.852 / .928
Stage 1 sensitivity / specificity	.795 / .933
Diseased images rejected at Stage 1	297 of 1,449
Stage 2 macro-AUC on received images	.955
Exact group set / exact disease set	.732 / .687
Images: correct healthy / exact set / partial	942 / 744 / 213
Images: total miss / false alarm	492 / 63

TABLE V: Three-stage cascade on the test split (

n=2{,}454

), all upstream errors propagated.

Fig. 2: Loss cascade: the stage at which each true disease instance of the test split is lost. The two coarse gates account for 34.2% of all instances.

Image	Truth	Cascade	Flat
4719_right	DR	exits at Stage 1 ( $s_{1}{=}0.24$ ): DR $\times$	DR $\times$
rfmid_1_20	MHL, TSLN, LS	G2 not routed: MHL $\times$ ; TSLN $\times$ at Stage 3; LS ✓; DR FP	LS ✓; MHL $\times$ ; TSLN $\times$
3903_left	healthy	$s_{1}{=}0.61$ , routed to G3: GLA FP	GLA FP
rfmid_1_240	TSLN, MH	TSLN ✓, MH ✓	TSLN ✓, MH ✓
2322_right	CAT	CAT ✓	CAT ✓

TABLE VI: Case studies on test images (✓ found,

\times

missed, FP false positive).

	Configuration	Exact	Jaccard	micro-F1	macro-F1	F1 well	macro-AUC	macro-AP	Detection
1	Flat baseline	.6789	.7229	.6642	.4426	.5767	.9430	.5048	.6207
2	Flat + TTA $\times$ 4	.6817	.7248	.6637	.4420	.5707	.9470	.5144	.6056
3	Flat + TTA $\times$ 4 + re-tuned thresholds	.6764	.7273	.6653	.4444	.5748	.9470	.5144	.6492
4	Three-stage cascade	.6870	.7278	.6495	.3917	.5248	.7467 ∗	.3713 ∗	.5939
5	Cascade + flat ensemble ( $w=0.5$ )	.6817	.7248	.6543	.4154	.5615	.9500	.5213	.5934
6	Label-correlation	.6679	.7239	.6577	.4518	.5573	.9662	.4599	.6818

TABLE VII: Ablation on the test split (2,454 images, 39 targets). All thresholds and blend weights are selected on validation. F1 well : macro-F1 over the 23 targets with

\geq 10

test positives. Best value per column in bold.

Fig. 3: Ablation at a glance (values of Table VII ). The shaded row is the best configuration; the label-correlation rows carry the highest detection.

Fig. 4: Test F1 for the 16 most frequent targets (number of test positives under each code).

Target	$n$	Cascade	Flat	Best	AP	Thr.	Target	$n$	Cascade	Flat	Best	AP	Thr.
DR	630	.762	.778	.785	.880	.70	RT	13	.316	.375	.421	.367	.35
ARMD	112	.594	.627	.650	.657	.40	CRVO	11	.632	.667	.667	.669	.72
MH	111	.764	.745	.745	.854	.90	CME	10	.286	.444	.588	.636	–
MYA	105	.802	.812	.838	.902	.68	TV	8	.000	.182	.235	.276	–
ODC	98	.532	.571	.627	.687	.80	CWS	8	.154	.400	.500	.363	–
CAT	95	.856	.829	.840	.922	.55	TD	8	.375	.182	.154	.211	–

TABLE VIII: Per-target test results, ordered by number of test positives (

n

). F1 of the three-stage cascade (native decisions), the flat model and the best configuration (flat + label-correlation ensemble + TTA); AP and decision threshold of the best configuration (“–”: default 0.5, fewer than ten validation positives). Best F1 per target in bold.

		Flat		Three-stage cascade
Source	$n$	macro-F1	Exact	macro-F1	Exact	Detection
ODIR-5K	1,642	.625	.746	.614	.757	.611
RFMiD 1.0	640	.375	.566	.335	.567	.602
RFMiD 2.0	172	.319	.459	.272	.465	.447

TABLE IX: Test performance by source dataset.

Study	Data	$n$	macro-F1	micro-F1	macro-AP
C-Tran [ 8 ]	MuReD	15	.657 / .638	–	.704 / .685 a
Dual CNN [ 12 ]	MuReD	15	.657 / .403	.726 / .548	.704 / .582 b
ASL cascade [ 13 ]	RFMiD	26	.478 / .476	.702 / .626	.530 / .539
12 models [ 11 ]	RFMiD	26	.478 / .340	.702 / .509	–
a Averaged over the comparator’s full label set. b Micro PR-AUC. For [ 11 ] , the best of twelve models per metric.

TABLE X: Subset-matched comparison (ours / theirs).

n

: shared classes.

Class	$n$	Ours	C-Tran	$\Delta$ F1
DR	630	.785	.859	$-$ 0.074
MH	111	.745	.724	$+$ 0.021
ODC	98	.627	.703	$-$ 0.076
TSLN	61	.646	.787	$-$ 0.141
ARMD	112	.650	.615	$+$ 0.035
DN	47	.549	.607	$-$ 0.058

TABLE XI: Per-class F1 against C-Tran [ 8 ] on the 15 shared classes (

n

: our test positives).

Multi-Label Classification

Multi-label classification assigns several potentially related tags to each example, from medical findings and topics to objects and user interests.

See all 134 papers

Less is More in Semantic Space: Intrinsic Decoupling via Clifford-M for Fundus Image Classification
Mar 21, 2026Yifeng ZhengMedical Image Classification Fundus Imaging
Retinal Disease Classification from Fundus Images using CNN Transfer Learning
Mar 24, 2026Ali AkramRetinal Disease Classification Medical Image Classification
CliftonLarsonAllen (CLA)
Explainable Convolutional Neural Networks for Retinal Fundus Classification and Cutting-Edge Segmentation Models for Retinal Blood Vessels from Fundus Images
May 12, 2024Fatema Tuj Johora Faria, Mukaffi Bin Moin, Pronay Debnath +2Explainable Medical Image Analysis Retinal Vessel Segmentation
aDepartment of Computer Science and Engineering, Ahsanullah University of Science and Technology, Dhaka, Bangladesh

Mar 21, 2026J/K move · Enter open · S save

Less is More in Semantic Space: Intrinsic Decoupling via Clifford-M for Fundus Image Classification

Yifeng Zheng

Medical Image Classification Fundus Imaging Multi-Scale Feature Fusion

Open paper AskPDF arXiv

Multi-label fundus diagnosis requires features that capture both fine-grained lesions and large-scale retinal structure. Many multi-scale medical vision models address this challenge through explicit frequency decomposition, but our ablation studies show that such heuristics provide limited benefit in this setting: replacing the proposed simple dual-resolution stem with Octave Convolution increased parameters by 35% and computation by a 2.23-fold increase in computation; without improving mean accuracy, while a fixed wavelet-based variant performed substantially worse. Motivated by these findings, we propose Clifford-M, a lightweight backbone that replaces both feed-forward expansion and frequency-splitting modules with sparse geometric interaction. The model is built on a Clifford-style rolling product that jointly captures alignment and structural variation with linear complexity, enabling efficient cross-scale fusion and self-refinement in a compact dual-resolution architecture. Without pre-training, Clifford-M achieves a mean AUC-ROC of 0.8142 and a mean macro-F1 (optimal threshold) of 0.5481 on ODIR-5K using only 0.85M parameters, outperforming substantially larger mid-scale CNN baselines under the same training protocol. When evaluated on RFMiD without fine-tuning, it attains 0.7425 +/- 0.0198 macro AUC and 0.7610 +/- 0.0344 micro AUC, indicating reasonable robustness to cross-dataset shift. These results suggest that competitive and efficient fundus diagnosis can be achieved without explicit frequency engineering, provided that the core feature interaction is designed to capture multi-scale structure directly.

Abstract

Multi-Label Classification

Explore similar work

Less is More in Semantic Space: Intrinsic Decoupling via Clifford-M for Fundus Image Classification

Retinal Disease Classification from Fundus Images using CNN Transfer Learning

Explainable Convolutional Neural Networks for Retinal Fundus Classification and Cutting-Edge Segmentation Models for Retinal Blood Vessels from Fundus Images