Retinal disease monitoring with optical coherence tomography angiography (OCTA) relies on segmentation of vessels and the foveal avascular zone (FAZ). Promptable models such as SEEM (Segment Everything Everywhere All at Once) allow click-based refinement, but without adaptation SEEM segments thin vessels poorly and needs many clicks. Fully fine-tuning its 342 million parameters on a few hundred images is costly and risks overfitting, and parameter-efficient methods such as low-rank adaptation (LoRA) usually adapt only the encoder. Yet SEEM reads every candidate mask from one feature map at quarter resolution, too coarse for vessels one to three pixels wide. This study introduces the Read-out Basis Adapter (RBA), whose read-out branch generates 32 image-specific basis maps at half resolution. Each candidate accesses them through its own vector and a zero-initialised projection, so the model starts exactly as SEEM. The encoder branch, a standard rank-1 LoRA, supplies the information needed to tell arteries from veins. RBA adds 0.47 million trainable parameters. Under simulated error-driven clicks, with the ground truth used only to place them, RBA outperforms LoRA, Conv-LoRA andAdaptFormer on OCTA-500 3mmlarge vessels. Averaged over three seeds, it reaches an intersection over union (IoU) of 85.16% and a Dice score of 91.93%, 2.90 and 1.72 percentage points above Conv-LoRA. On the 6 mm subset, one run gives margins of 4.35 and 2.76 points. The gain extends to arteries, veins and ROSE-O but not to the FAZ. In SEEM, adaptation capacity pays more at the mask read-out than in the encoder.
Figures & tables
Figure 1: Three ways to adapt a promptable segmentation model: (a) encoder adapters such as LoRA, Conv-LoRA and AdaptFormer, (b) output-token refinement as in HQ-SEEM, and (c) RBA, which adapts the read-out basis of SEEM. Green marks trainable additions.
Figure 2: Overview of SEEM with RBA. Blue marks the image path, purple the click path and green the read-out term of RBA. Dashed loops mark the next click and the previous-mask memory.
Figure 3: The RBA module: (a) read-out branch and (b) encoder branch. Green numbers give parameter counts, dashed green boxes mark zero initialisation, and red marks the unchanged read-out of SEEM.
Method
Enc.
Read-out
Params
Dice@4
clDice
SE
SP
NoC@90
RBA − row [95% CI]
RBA better
No adapter
–
–
0
75.83
80.44
80.05
97.79
20.00
–
–
LoRA r=1
✓
–
0.27 M
89.47
92.34
90.27
99.18
11.80
+2.48 [ +2.35 , +2.61 ]
50/50
LoRA r=16‡
✓
–
4.33 M
90.13 ± 0.04
92.80 ± 0.02
90.79 ± 0.14
99.24 ± 0.02
7.37 ± 0.53
+1.82 [ +1.71 , +1.94 ]
50/50
Conv-LoRA r=16‡
✓
–
4.35 M
90.22 ± 0.02
92.90 ± 0.02
90.96 ± 0.12
99.24 ± 0.01
6.68 ± 0.38
+1.73 [ +1.63 , +1.83 ]
50/50
AdaptFormer
✓
–
0.45 M
89.52
92.53
90.68
99.15
11.18
+2.44 [ +2.28 , +2.59 ]
50/50
Read-out only (ours)
–
✓
0.20 M
90.56
92.15
93.27
99.10
5.64
+1.40 [ +1.18 , +1.61 ]
50/50
Table 1: Adapters on large vessels, OCTA-500 3 mm subset, 50 test images, scores in %. Enc. and Read-out mark the two branches. Rows marked ‡ give the mean ± standard deviation over three training seeds; the other rows use seed 42. The last two columns give the paired change from the row to RBA for seed 42, with its 95% interval, and the number of images on which RBA is better; p<10−10 except † ( p=0.012 ). Bold: best; underlined: second best.
Figure 4: Dice@4 on large vessels against adapter parameters on a logarithmic scale, for seed 42. Blue marks encoder-only adapters, green marks adapters with the RBA read-out, and dashed arrows mark its addition.
Decoder trained
SEEM frozen
Adapter
Trained
Dice@4
Trained
Dice@4
Change
LoRA r=1
40.65
89.47
0.84
89.44
−0.03
Conv-LoRA r=16
44.73
90.23
4.91
90.26
+0.03
RBA
40.85
91.95
1.04
91.94
−0.01
RBA, click prompt off
–
–
0.47
91.75
–
Table 2: Effect of freezing the decoder of SEEM on large vessels (seed 42). Trained gives millions of trained parameters, and Change is frozen minus trained, in %. In the last row the click prompt is also off.
Dice@4
Dataset
Target
RBA
Conv-LoRA
Δ [95% CI]
Better
p
clDice R / C
NoC@90 R / C
OCTA-500 3 mm
Large vessels ‡
91.93
90.22
+1.73 [ +1.63 , +1.83 ]
50/50
<10−10
93.88 / 92.90
3.03 / 6.68
Veins
90.49
88.87
+1.62 [ +1.31 , +1.92 ]
45/50
<10−10
92.78 / 92.28
6.72 / 13.54
Arteries
91.26
89.69
+1.57 [ +1.30 , +1.84 ]
48/50
<10−10
93.17 / 92.39
4.48 / 10.48
FAZ
97.52
97.86
−0.34 [ −0.90 , +0.21 ]
25/50
0.59
–
1.00 / 1.00
OCTA-500 6 mm
Large vessels
88.67
85.91
+2.76 [ +2.59 , +2.92 ]
50/50
<10−10
91.08 / 89.44
13.40 / 20.00
Table 3: RBA against Conv-LoRA r=16 on every target, scores in %. Δ is the paired change in Dice@4 with its 95% interval for seed 42, and p is Holm-corrected. ‡ Mean over three seeds. Bold marks the better Dice@4 where the difference is significant.
Dice
IoU
ACC
SE
SP
Dataset
Target
C
R
C
R
C
R
C
R
C
R
OCTA-500 3 mm
Large vessels ‡
90.22
91.93
82.26
85.16
98.69
98.92
90.96
92.52
99.24
99.38
Veins
88.87
90.49
80.03
82.69
99.35
99.44
90.36
91.24
99.62
99.69
Arteries
89.69
91.26
81.39
84.00
99.23
99.35
90.52
92.02
99.57
99.63
FAZ
97.86
97.52
95.88
95.32
99.84
99.80
98.58
98.89
99.89
99.83
OCTA-500 6 mm
Large vessels
85.91
88.67
75.39
79.74
97.25
97.82
88.45
89.70
98.18
98.68
Table 4: Pixel measures of Conv-LoRA r=16 (C) and RBA (R) after four clicks, in %. ‡ Mean over three seeds; the other rows use seed 42. Bold marks the better of the two.
Large vessels
Arteries
Veins
FAZ
Method
Setting
3 mm
6 mm
3 mm
6 mm
3 mm
6 mm
3 mm
6 mm
U-Net [ 37 ]
Automatic, six projections
92.04
89.11
89.58
87.27
87.99
86.90
96.48
87.55
Attention U-Net [ 35 ]
Automatic, six projections
91.91
89.26
90.14
87.35
88.07
87.03
97.76
86.73
IPN-V2 [ 26 ]
Automatic, OCTA volume
92.32
89.17
90.78
87.27
89.71
87.30
97.68
90.12
IPN V2+ [ 27 ]
Automatic
92.74
89.41
–
–
–
–
97.55
90.84
FARGO [ 36 ]
Automatic
91.68
89.15
–
–
–
–
98.39
92.72
Table 5: Published Dice scores on the OCTA-500 test sets in %, for context. Results of other methods are taken from the cited papers. A dash marks values not reported, † marks images 10200–10300 instead of 10251–10300, and ‡ marks the mean over three seeds. Bold: best; underlined: second best.
Method
Type
Vessels
FAZ
U-Net [ 37 ]
Single-task
70.32
93.82
OCTA-Net [ 32 ]
Single-task
75.83
94.64
UTNet [ 13 ]
Single-task
75.08
94.47
FARGO [ 36 ]
Multi-task
74.59
94.42
MTI-Net [ 43 ]
Multi-task
75.34
94.77
VAFF-Net [ 14 ]
Multi-task
76.58
95.19
Table 6: Published Dice scores on the ROSE-O test set in %, for context. Bold: best; underlined: second best.
Encoder branch
Read-out branch
Params (M)
Large vessels
Veins
–
–
0
75.83
50.98
✓
–
0.27
89.47
87.98
–
✓
0.20
90.56
71.32
✓
✓
0.47
91.95
90.49
Read-out added to encoder
+2.48
+2.51
Encoder added to read-out
+1.40
+19.16
Table 7: Ablation of the two branches of RBA, Dice@4 in % on 50 test images (seed 42), with paired changes and the interaction. The decoder of SEEM is trained in every row.
Row
Dice@4
Found
Art. as vein
FP on art.
No adapter
50.98
64.5
41.3
59.8
Read-out only
71.32
85.7
30.4
69.5
LoRA r=1
87.98
89.6
2.1
9.4
Conv-LoRA r=16
88.87
90.3
1.8
7.4
Conv-LoRA + read-out
89.68
93.0
2.4
12.1
RBA (ours)
90.49
91.2
2.3
15.1
Table 8: Vessel identity on the vein target, in % (seed 42). Found is the share of vein pixels found, Art. as vein the share of artery pixels marked as vein, and FP on art. the share of false positives on arteries.
Row
Dice@4
Recall, thin
Recall, mid
Recall, thick
Edge error
Centre-line recall
Distant FP
No adapter
75.83
52.5
88.7
94.9
24.8
76.8
18.5
LoRA r=1
89.47
77.6
94.3
97.2
11.7
90.3
9.4
Conv-LoRA r=16
90.23
79.5
94.5
97.4
10.8
91.2
11.0
Read-out only
90.56
85.8
95.6
97.5
9.0
94.1
29.3
Conv-LoRA + read-out
91.82
84.2
95.1
97.7
8.6
93.5
17.8
RBA (ours)
91.95
84.0
94.9
97.4
8.4
93.5
19.2
Table 9: Large-vessel results by vessel width and edge proximity, in % (seed 42). Lower is better for the edge error and for distant false positives (FP). The last rows give the paired change when the read-out is added, with the number of images that improved. Bold marks the best value.
Figure 5: Large-vessel segmentation after four clicks on test image 10461, with the largest gain, and image 10498, with the median gain. White marks correct vessel, blue missed vessel and orange false vessel. Numbers are Dice scores.
Figure 6: Vein and artery segmentation after four clicks on the images with the largest and the median gain. Colours as in Fig. 5 , with pink for false vessel on the other vessel type and grey for the other vessel type in the ground truth. Numbers are Dice scores.
Method
Trained (M)
Memory (GB)
Encoder (ms)
Click (ms)
Image (ms)
No adapter
40.38
1.70
92
74.6
390
LoRA r=1
40.65
1.70
103
72.3
393
LoRA r=16
44.71
1.71
105
72.3
394
Conv-LoRA r=16
44.73
1.71
111
72.2
400
AdaptFormer
40.83
1.70
94
71.6
381
Read-out only
40.58
1.78
99
72.3
389
Table 10: Cost at test time on the 3 mm subset, large vessels, on one NVIDIA GeForce RTX 3070 Ti Laptop GPU. Encoder is the pass run once per image, and Image adds four click rounds. Trained is in millions of parameters, and Memory is peak GPU memory.
Interactive medical image segmentation uses clinician prompts and iterative model feedback to refine anatomical masks and reduce manual annotation effort.
Retinal vessel segmentation in multi-domain deployment requires a source model to adapt to domains that differ in imaging conditions and annotation conventions. Conventional parameter-efficient fine-tuning reduces target-specific storage, but its highly restricted adaptation subspace can be insufficient for reconstructing thin, connected vascular structures. We therefore ask how target-specific capacity should be allocated so that topology-aware supervision remains effective under a strict per-domain parameter budget. Based on this principle, we propose TAPDecoderFT, a topology-responsive, role-structured adaptation framework. Specifically, TAPDecoderFT shares a fixed source parameter state across deployment domains, uses low-rank residuals for target-specific private/fusion feature mixing, and retains a trainable dense-reconstruction path comprising the decoder, output head, and refinement module. To promote structurally faithful predictions, the compact target state is jointly optimized with a region-overlap and topology-aware objective that encourages centerline continuity and thin-branch recovery. It improves both DSC and clDice over GenericLoRA-r4 and narrow TAP-r4 in all six directions and is comparable to full fine-tuning.
Existing deep learning frameworks for Optical Coherence Tomography Angiography (OCTA) vessel segmentation are largely derived from the U-Net architecture, which serves as the foundation for most current designs. However, most of these methods focus only on holistic representation, struggling to address the problem of low local contrast unique to OCTA, which leads to vessel discontinuities and loss of detail. To address these problems, we propose LSENet, which builds upon the U-Net architecture by introducing three core innovative modules: To address vessel discontinuities, we introduce the Patch Information Enhance module (PIE), which replaces standard skip connections to execute patch-wise attention. To mitigate detail loss, the Multiscale Feature Fusion module (MFF) is proposed to feed the PIE module rich, multi-scale information by extracting visually interpretable features from both the original input and preceding layers. Finally, the Connectivity Refinement Decoder (CRD) is designed to refine features from all levels and utilize a large kernel in the final convolutional layer to reduce fragmentation. Experiments on three public datasets (OCTA-500, ROSE-1, and ROSSA) demonstrate that our proposed LSENet achieves state-of-the-art performance while requiring fewer parameters.
Tuopusen Huang, Ding Ma, Xiangqian Wu
Faculty of Computing Harbin Institute of Technology Harbin, China
Ultra-widefield (UWF) swept-source optical coherence tomography angiography (SS-OCTA) enables large-area retinal vascular imaging, yet vessel segmentation at this scale lacks dedicated public benchmarks and comprehensive evaluation for quantitative vascular analysis. We introduce WOIVES, to our knowledge the first publicly available UWF SS-OCTA vessel-segmentation dataset, comprising 206 eyes from 152 participants with a 24x20mm^2 field of view. WOIVES spans emmetropia to high myopia and provides soft probability vessel annotations. We further propose PG-Mamba, a visual state space model that enhances conventional directional scans with two complementary polar-coordinate scan orders. An auxiliary Dynamic FOV Gating module performs spatial modulation at the bottleneck. PG-Mamba outperformed seven competitive approaches on broad segmentation metrics under cross-validation. It achieved the lowest median absolute errors for vessel density, fractal dimension, and vessel length density. WOIVES is publicly available on Zenodo (DOI: 10.5281/zenodo.21904672), and the PG-Mamba code is available at https://github.com/syb1234567/PG-Mamba.
Yang Liu, Yibing Shen, Keming Zhao +12
Tsinghua Shenzhen International Graduate School, Shenzhen 518109, China · Tsinghua University, Beijing, China · Medical Artificial Intelligence Lab, The First Hospital of Hebei Medical University, Hebei Medical University, Shijiazhuang 050000, China +2