TRIAGE: Risk-Controlled Pseudo-Label Admission for Annotation-Efficient Semi-Supervised Retinal OCT Classification

Md Ashraful Hossen Akash; Shyla Afroge; Abdullah Al Mamun; Md. Kishor Morol; Tze Hui Liew

cs.CVAug 14, 2026

TRIAGE: Risk-Controlled Pseudo-Label Admission for Annotation-Efficient Semi-Supervised Retinal OCT Classification

Authors: Md Ashraful Hossen Akash, Shyla Afroge, Abdullah Al Mamun, Md. Kishor Morol, Tze Hui Liew

Organizations: Department of Computer Science & Engineering, Rajshahi University of Engineering & Technology, Rajshahi-6204, Bangladesh · Elite Research Lab LLC, New York, USA · Faculty of Information Science and Technology, Multimedia University, Melaka, Malaysia · Centre for Intelligent Cloud Computing (CICC), COE of Advanced Cloud, Faculty of Information Science & Technology, Multimedia University, 75450, Melaka, Malaysia

Abstract

The advanced retinal disease diagnosing imaging modality, optical coherence tomography (OCT), encounters a lack of automation because of the high expenses for annotations performed by specialists. The use of SSL solves the problem of insufficient annotations using unlabeled B-scans; however, most of the current techniques for generating pseudo-labels are based on prediction confidence without considering the asymmetry between different types of errors. This paper proposes TRIAGE, a risk-controlled semi-supervised framework for OCT scans classification, which uses the concept of a patient-level conformal risk controller with an asymmetric cost matrix. TRIAGE unites three crucial modules: a hierarchical classifier that is capable of working with partially abnormal supervision of the disease subtypes, a patient-grouped conformal risk controller with primal-dual coverage control, and a context-aware Transformer teacher for cross-slice verification. On the dataset from Noor Eye Hospital (16,822 B-scans, 161 patients, and 554 volumes) with a test set of unseen patients, TRIAGE demonstrates 89.66% scan-level accuracy, 0.8805 macro-F1, 0.9641 macro-AUC, and an 8.34% under-grading rate when using only 20% of the labeled data. With only 5% of the labeled data, TRIAGE keeps 76.88% accuracy and a 0.1656 under-grading rate. Compared with the other six state-of-the-art semi-supervised methods, TRIAGE significantly outperforms them with ablation study demonstrating the contribution of each module in the overall framework performance (by 42.7% in terms of under-grading rate comparing to fixed threshold methods). TRIAGE demonstrates 98.00% accuracy for 3-class classification with 1% labeled data and 95.94% accuracy for 8-class classification with 10% labeled data on the OCT-C8 dataset.

Figures & tables

Figure 1: Overall abstract architecture and data pipeline of TRIAGE, illustrating patient-disjoint splitting, weak/strong view generation, training loop, and dual-level evaluation.

Figure 2: Detailed main model architecture and dual-teacher workflow of TRIAGE, illustrating single-scan risk gating, volume-context Transformer teacher agreement, and student supervision.

Panel of Figure 3 — Figure 3: Hierarchical backbone: (a) Feature extraction pipeline with dual classification heads and CoMatch projection head, and (b) 7-stage EfficientNet-B0 backbone with internal MBConv block operations including Squeeze-and-Excitation attention and Stochastic Depth.

True $\backslash$ Predicted	normal	drusen	cnv
normal	0	1	1
drusen	2	0	1
cnv	5	2	0

Table 1: Clinical cost matrix

R[\text{true}][\text{predicted}]

for the Noor 3-class taxonomy.

Panel of Figure 4 — Figure 4: Conformal risk control: (a) Patient-grouped calibration pipeline, and (b) Three-outcome admission rule decision tree.

Panel of Figure 5 — Figure 5: 3D Volume processing: (a) Transformer context encoder across adjacent B-scans, and (b) Dual-teacher agreement gating and risk-sensitive volume pooling.

Figure 6: Representative B-scan samples from Noor Eye Hospital: normal , drusen , and cnv .

Split	Patients	Volumes	Images	normal	drusen	cnv
Labelled (20%)	21	78	2,329	1,233	676	420
Unlabelled	87	296	9,012	4,523	2,634	1,855
Validation	6	19	563	317	114	132
Calibration	14	45	1,406	750	443	213
Test	33	116	3,512	1,761	1,131	620
Total	161	554	16,822	8,584	4,998	3,240

Table 2: Noor dataset patient-disjoint splits.

Panel of Figure 7 — Figure 7: Representative image samples from OCT-C8 across (a) 3-class and (b) 8-class taxonomies.

(a) Scan level, per class
Class	Precision	Recall	F1	Support
normal	0.9206	0.9614	0.9406	1,761
drusen	0.9103	0.7984	0.8507	1,131
cnv	0.8120	0.8919	0.8501	620
Macro average	0.8810	0.8839	0.8805	3,512
Weighted average	0.8981	0.8966	0.8956	3,512

Table 3: Test performance of TRIAGE on Noor at a 20% label budget.

Panel of Figure 8 — Figure 8: Diagnostic performance of TRIAGE on the Noor dataset at a 20% label budget: (a) Scan-level confusion matrix, (b) One-vs-Rest ROC curves, and (c) Precision–Recall curves across diagnostic classes.

Diagnostic	Value
Pseudo-label coverage (epoch-9 mean over training batches)	0.574
Pseudo-label accuracy on fully admitted labels	0.837
Partial-label rate (fraction of admissions that are partial)	0.036
Single-scan / volume-context teacher agreement rate	0.687
Calibrated risk threshold $\lambda$	0.050
Dual coverage-relaxation variable $\mu$	0.008

Table 4: Training diagnostics at the selected epoch-9 checkpoint, Noor dataset, 20% labelled.

Interval (epoch)	Risk threshold $\lambda$	Coverage	Partial-label rate	Teacher agreement
1	0.050	0.107	0.000	0.512
2	0.099	0.521	0.002	0.598
3	0.099	0.548	0.008	0.621
4	0.139	0.562	0.014	0.643
5	0.139	0.589	0.019	0.655
6	0.099	0.571	0.023	0.662

Table 5: Calibrated risk threshold

\lambda

, pseudo-label coverage, partial-label rate and teacher agreement across the ten evaluation intervals, Noor dataset, 20% labelled.

(a) Scan-level evaluation
Method	Accuracy	F1	AUC	UGR
Supervised (20%)	79.25 $\pm$ 0.85	78.70 $\pm$ 0.92	92.31 $\pm$ 0.65	15.68 $\pm$ 1.20
Mean Teacher [ 10 ]	82.80 $\pm$ 0.80	82.92 $\pm$ 1.00	94.42 $\pm$ 0.50	10.17 $\pm$ 1.10
Pseudo-Label [ 9 ]	86.70 $\pm$ 0.75	85.04 $\pm$ 0.85	95.09 $\pm$ 0.55	11.94 $\pm$ 1.30
FixMatch [ 11 ]	87.72 $\pm$ 0.70	86.90 $\pm$ 0.80	94.92 $\pm$ 0.60	14.56 $\pm$ 1.40
FreeMatch [ 14 ]	87.87 $\pm$ 0.65	86.48 $\pm$ 0.75	96.25 $\pm$ 0.50	13.93 $\pm$ 1.25

Table 6: Comparison against SSL baselines and supervised references on Noor dataset under a 20% label budget. Results represent mean

\pm

standard deviation across three random seeds (12, 111, 165). Best SSL performance is in bold.

(a) Scan-level evaluation
Method	Accuracy	F1	AUC	UGR
Supervised (5%)	60.28 $\pm$ 1.60	57.50 $\pm$ 1.95	76.00 $\pm$ 1.40	50.03 $\pm$ 2.50
Mean Teacher [ 10 ]	67.20 $\pm$ 1.40	66.32 $\pm$ 1.60	82.94 $\pm$ 1.10	21.07 $\pm$ 1.90
Pseudo-Label [ 9 ]	67.85 $\pm$ 1.50	58.99 $\pm$ 2.10	87.25 $\pm$ 1.00	52.88 $\pm$ 2.80
FreeMatch [ 14 ]	67.03 $\pm$ 1.50	49.96 $\pm$ 2.20	84.14 $\pm$ 1.20	19.25 $\pm$ 2.10
FlexMatch [ 13 ]	71.36 $\pm$ 1.50	69.19 $\pm$ 1.20	86.52 $\pm$ 1.00	33.24 $\pm$ 2.40

Table 7: Comparison against SSL baselines and supervised references on Noor dataset under a 5% label budget. Results represent mean

\pm

standard deviation across three random seeds (12, 111, 165). Best SSL performance is in bold.

Panel of Figure 9 — Figure 9: Diagnostic performance of TRIAGE on the Noor dataset at a 5% label budget: (a) Scan-level confusion matrix, (b) One-vs-Rest ROC curves, and (c) Precision–Recall curves across diagnostic classes.

Label budget	Accuracy	Macro-F1	Macro-AUC	UGR
5%	0.7688	0.7158	0.8833	0.1656
20%	0.8966	0.8805	0.9641	0.0834

Table 8: Label-efficiency comparison for TRIAGE, Noor dataset, scan level.

Configuration	Accuracy	F1	AUC	UGR
Fixed threshold (0.95)	87.72 $\pm$ 0.45	85.90 $\pm$ 0.60	93.92 $\pm$ 0.35	14.56 $\pm$ 0.90
Class-adaptive threshold	88.35 $\pm$ 0.40	86.60 $\pm$ 0.55	94.85 $\pm$ 0.40	12.30 $\pm$ 0.85
w/o volume teacher	89.02 $\pm$ 0.35	87.40 $\pm$ 0.45	95.00 $\pm$ 0.30	9.75 $\pm$ 0.70
w/o conformal rule	89.28 $\pm$ 0.32	87.67 $\pm$ 0.42	95.55 $\pm$ 0.28	9.10 $\pm$ 0.65
TRIAGE (full)	89.66 $\pm$ 0.32	88.05 $\pm$ 0.41	96.41 $\pm$ 0.25	8.34 $\pm$ 0.55

Table 9: Component ablation on Noor dataset under a 20% label budget. Results represent mean

\pm

standard deviation across three random seeds (12, 111, 165). Best performance is in bold.

Comparison	Metric	Difference	Test	$p$ -value
TRIAGE vs FlexMatch (20%)	Accuracy	+0.54%	Paired $t$ -test	0.041
TRIAGE vs FlexMatch (20%)	Macro-F1	+0.37%	Paired $t$ -test	0.028
TRIAGE vs CoMatch (20%)	UGR	-4.00%	Wilcoxon test	0.012
TRIAGE vs FlexMatch (5%)	Accuracy	+5.52%	Paired $t$ -test	$<0.005$
TRIAGE vs CoMatch (5%)	UGR	-26.62%	Wilcoxon test	$<0.001$

Table 10: Statistical significance analysis comparing TRIAGE against leading baselines across three random seeds (12, 111, 165).

Seed	Accuracy	F1	AUC	UGR
12	89.42	87.64	96.21	8.91
111	89.88	88.31	96.62	7.82
165	89.68	88.20	96.40	8.29
Mean	89.66	88.05	96.41	8.34
Std	0.32	0.36	0.21	0.55

Table 11: Run-wise performance metrics of TRIAGE across three independent random seeds (12, 111, 165) on Noor dataset (20% budget).

Dataset	Classes	Labels	Accuracy	Macro-F1	Macro-AUC	UGR
OCT-C8	3	1%	0.9800	0.9795	0.9977	0.0179
OCT-C8	8	10%	0.9594	0.9598	0.9969	0.0299
Noor	3	5%	0.7688	0.7158	0.8833	0.1656
Noor	3	20%	0.8966	0.8805	0.9641	0.0834

Table 12: Cross-dataset generalisation of TRIAGE, scan level. The volume teacher is disabled for all OCT-C8 rows. Noor rows are repeated for reference only and are not directly comparable, as they use a different taxonomy, label budget and volume structure.

Panel of Figure 10 — Figure 10: Diagnostic performance of TRIAGE on OCT-C8 under the three-class taxonomy at a 1% label budget: (a) Confusion matrix, (b) One-vs-Rest ROC curves, and (c) Precision–Recall curves across diagnostic classes.

Panel of Figure 11 — Figure 11: Diagnostic performance of TRIAGE on OCT-C8 under the eight-class taxonomy at a 10% label budget: (a) Confusion matrix, (b) One-vs-Rest ROC curves, and (c) Precision–Recall curves across diagnostic classes.

(a) OCT-C8 three-class subset ( normal / amd / dme )
Method	Labelled budget	Test images	Accuracy
Li et al. (class-aware SSL) [ 37 ]	1.1% (102/9,000)	1,050 ∗	0.9640
TRIAGE (ours)	1.0% (90/9,000)	1,800	0.9800

Table 13: Comparison against published results on the same image collections.

Figure 12: Visual model interpretability via LayerCAM across retinal OCT categories. For each category, the original input B-scan (top), the high-resolution LayerCAM activation heatmap (middle), and the blended visual overlay (bottom) demonstrate that model attention precisely aligns with anatomical retinal biomarkers (e.g., subretinal fluid in CNV/DME, RPE elevations in Drusen, and foveal layer disruptions in MH), while maintaining diffuse, low-intensity attention on normal scans.

Explore similar work

PC-Seg: Progressive Cross-View Consistency for 3D OCT Segmentation from Sparse 2D Annotations
Jul 20, 2026Tsubasa Konno, Takahiro Ninomiya, Yukun Zhou +6Multi-View Consistency Optical Coherence Tomography
Graduate School of Information Sciences, Tohoku University, Japan · Department of Ophthalmology, Graduate School of Medicine, Tohoku University · Institute of Ophthalmology, University College London, UK +3
Anatomy-Aware Unsupervised Detection and Localization of Retinal Abnormalities in Optical Coherence Tomography
Apr 24, 2026Tania Haghighi, Sina Gholami, Hamed Tabkhi +1Medical Imaging Medical Image Anomaly Detection
University of North Carolina at Charlotte Charlotte, NC, USA
Gradient based Severity Labeling for Biomarker Classification in OCT
Feb 23, 2026Kiran Kokilepersaud, Mohit Prabhushankar, Ghassan AlRegib +2Classification Pseudo-Labeling
OLIVES at the Center for Signal and Information Processing CSIP, School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA · Retina Consultants Texas, Retina Consultants of America, Houston, Texas, USA

Jul 20, 2026J/K move · Enter open · S save

PC-Seg: Progressive Cross-View Consistency for 3D OCT Segmentation from Sparse 2D Annotations

Tsubasa Konno, Takahiro Ninomiya, Yukun Zhou, Koichi Ito+5

Graduate School of Information Sciences, Tohoku University, Japan · Department of Ophthalmology, Graduate School of Medicine, Tohoku University · Institute of Ophthalmology, University College London, UK+3

Multi-View Consistency Optical Coherence Tomography Semi-Supervised Medical Image Segmentation Medical Image Segmentation 3D Medical Image Segmentation

Open paper AskPDF arXiv

Volumetric segmentation of optical coherence tomography (OCT) images is essential for diagnosing ocular diseases but requires labor-intensive voxel-wise annotations. While semi-supervised learning (SSL) can reduce annotation costs, most existing methods process data slice by slice and fail to exploit the inherent 3D spatial context. We propose PC-Seg, a progressive cross-view consistency framework that learns high-accuracy 3D segmentation models from sparse 2D annotations. Unlike conventional multi-view approaches, PC-Seg uses a single 2D model to learn cross-view consistency from standard B-scans and orthogonal slices, thereby generating reliable volumetric pseudo-labels. These pseudo-labels are then distilled into a 3D model, followed by a co-training stage in which the 2D and 3D models mutually refine each other through ensemble pseudo-labeling. Experiments on the MSHC and Duke DME datasets demonstrate that PC-Seg achieves accuracy comparable to fully supervised learning while using labels for only about 0.7% of the training data, outperforming state-of-the-art semi-supervised and retinal layer segmentation methods. Our code is publicly available at https://github.com/gsisaoki/pc-seg-official.

Abstract

Explore similar work

PC-Seg: Progressive Cross-View Consistency for 3D OCT Segmentation from Sparse 2D Annotations

Anatomy-Aware Unsupervised Detection and Localization of Retinal Abnormalities in Optical Coherence Tomography

Gradient based Severity Labeling for Biomarker Classification in OCT