Evaluating SAM-based and Convolutional Architectures in Thin Retinal Vessel Segmentation: Model Capacity, Computational Cost and Foundation Pre-training

Carlos Eduardo Fontaneli; Cesar Henrique Comin

cs.CVOct 30, 2026

Evaluating SAM-based and Convolutional Architectures in Thin Retinal Vessel Segmentation: Model Capacity, Computational Cost and Foundation Pre-training

Authors: Carlos Eduardo Fontaneli, Cesar Henrique Comin

Organizations: Department of Computer Science, Federal University of S˜ao Carlos, Rod. Washington Lu´ıs, km 235, 13565-905, S˜ao Carlos, SP, Brazil

Abstract

In retinal vessel segmentation, plain Vision Transformers (ViTs) are commonly expected to perform worse than convolutional neural networks (CNNs) on thin vessels, for reasons such as their coarse patch tokenization, their weak locality bias and their need for large training sets. Although this expectation is widely shared in the literature, how plain ViTs actually compare with CNNs on thin vessels has not been systematically quantified. We address this question by comparing a model with a plain ViT encoder (MedSAM), a hybrid Transformer (LiteMedSAM) and three CNNs of different sizes on three retinal datasets (DRIVE, HRF, FIVES) and, as a second-domain check, a microscopy dataset (VessMap). We measured the centerline recall and precision of thin vessels and compared the models while also considering their computational cost and how many epochs they needed to train. We found that the pre-trained MedSAM had a centerline recall of thin vessels that was not significantly different from that of the best CNN on all four datasets (-0.6 to +2.5 percentage points). A compact plain ViT obtained through architecture search but trained from scratch did not retain this performance. For pre-trained plain ViTs, our results do not support the expectation that they miss thin vessels, although similar thin-vessel performance can be obtained with much smaller models.

Figures & tables

Model	Type	Encoder	Params (M)
MedSAM	Plain ViT	ViT-B	93.7
LiteMedSAM	Hybrid	TinyViT	9.8
ConvNeXt	CNN	ConvNeXt-L	211.4
ConvUNeXt	CNN	ConvNeXt blocks	3.5 / 13.8
LW-Net	CNN	Conv. blocks	0.07

Table 1: The five architectures and their parameters. ConvUNeXt and LW-Net are U-Nets without a separate, named encoder, so for them the encoder column gives the type of block of the encoding path. The size of ConvUNeXt depends on the dataset (Appendix A ).

		CNN			Fine-tuned
Dataset	Metric	LW-Net	ConvUNeXt	ConvNeXt	LiteMedSAM	MedSAM
DRIVE ( $n=20$ )	Dice	$81.1\pm 2.4$	$77.6\pm 15.1$	$77.5\pm 2.5$	$82.3\pm 1.9$	$81.3\pm 1.3$
	$R_{\mathrm{cl}}$	$74.9\pm 7.2$	$70.1\pm 16.7$	$66.9\pm 7.5$	$79.4\pm 7.1$	$75.5\pm 6.9$
	$P_{\mathrm{cl}}$	$89.3\pm 2.3$	$90.5\pm 4.4$	$89.5\pm 3.1$	$87.6\pm 2.5$	$87.3\pm 2.9$
HRF ( $n=14$ )	Dice	$76.9\pm 6.9$	$79.7\pm 5.1$	$76.9\pm 6.2$	$80.6\pm 5.9$	$81.7\pm 4.5$
	$R_{\mathrm{cl}}$	$72.0\pm 12.0$	$72.8\pm 9.5$	$73.8\pm 7.6$	$82.2\pm 7.6$	$76.6\pm 8.6$

Table 2: Overall segmentation performance. Values in bold indicate the best performance. Underlined values are not significantly different from the best (bold) of the same row.

Figure 1: Mean centerline recall for different vessel radius ranges. The first range, radius in

[0.5,1.5)

px, shows

R_{\mathrm{thin}}

. Grey bars show the number of ground-truth centerline pixels for each range, calculated on the test images.

Figure 2: Centerline precision against centerline recall of thin vessels, per dataset.

Figure 3: Thin vessel detection agreement among the models. (a–d) Window of each test set with the most thin centerline pixels. Non-thin vessels are shown in dark grey. (e) Agreements between the models over all test images.

Figure 4: Cost of each architecture. (a) parameters and (b) GFLOPs per test image, log scale.

Figure 5: Relationship between performance and cost.

R_{\mathrm{thin}}

(a–d) and Dice (e–h) against GFLOPs per test image.

Figure 6: Effect of pre-training. (a) Dice, (b)

R_{\mathrm{thin}}

and (c)

P_{\mathrm{thin}}

of the two transformer-based models fine-tuned (filled) and trained from scratch (hollow), next to the CNNs.

Figure 7: Training epochs required to reach 95% of the best validation Dice. Models were validated every epoch or every 10 epochs (see Table A.1 for details). “

\leq

” means that the model reached the performance at the first validation.

		DRIVE ( $n=20$ )			VessMap ( $n=30$ )
MedSAM	Metric	Padded	Upscaled	$\Delta$	Padded	Upscaled	$\Delta$
Fine-tuned	Dice	$75.2\pm 1.8$	$81.3\pm 1.3$	$+6.1^{*}$	$78.7\pm 7.5$	$92.8\pm 2.1$	$+14.1^{*}$
	$R_{\mathrm{thin}}$	$39.0\pm 10.9$	$56.9\pm 11.8$	$+17.9^{*}$	$41.3\pm 12.7$	$82.6\pm 14.1$	$+41.3^{*}$
	$P_{\mathrm{thin}}$	$58.8\pm 7.9$	$73.0\pm 6.1$	$+14.2^{*}$	$76.8\pm 13.2$	$79.6\pm 14.8$	$+2.8$
From scratch	Dice	$65.7\pm 3.0$	$78.7\pm 1.6$	$+13.0^{*}$	$68.2\pm 6.4$	$89.7\pm 3.7$	$+21.5^{*}$
	$R_{\mathrm{thin}}$	$25.3\pm 8.0$	$39.3\pm 11.2$	$+14.0^{*}$	$43.6\pm 19.2$	$70.2\pm 20.9$	$+26.6^{*}$

Table 3: MedSAM evaluated with the padded and the upscaled protocol. Bold indicates the best value. Significant differences are indicated by an asterisk.

	Dice				$R_{\mathrm{thin}}$
Dataset	MedSAM		Best CNN		MedSAM		Best CNN
DRIVE	$81.3\pm{}$	$1.3$	$81.1\pm{}$	$2.4$ L	$56.9\pm{}$	$11.8$	$57.5\pm{}$	$12.1$ L
HRF	$81.7\pm{}$	$4.5^{*}$	$79.7\pm{}$	$5.1$ U	$26.9\pm{}$	$17.2$	$25.4\pm{}$	$13.5$ N
FIVES	$92.2\pm{}$	$8.5^{*}$	$90.2\pm{}$	$8.5$ N	$36.2\pm{}$	$11.9$	$36.5\pm{}$	$12.2$ N
VessMap	$92.8\pm{}$	$2.1$	$92.9\pm{}$	$2.1$ U	$82.6\pm{}$	$14.1$	$80.1\pm{}$	$12.2$ L

Table 4: Fine-tuned MedSAM against the best CNN of each metric and dataset. L: LW-Net , U: ConvUNeXt , N: ConvNeXt . Bold indicates the best value. Asterisks indicate significantly different values.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	Dataset	Learning rate	Weight decay	Batch size	Architecture	Val.	GPU h
LW-Net	DRIVE	$10^{-6}$ – $10^{-1}$	$10^{-6}$ – $10^{-1}$	2, 4, 8	first width 8, 16, 32	1	13.0
	HRF	$10^{-6}$ – $10^{-1}$	$10^{-6}$ – $10^{-1}$	2, 4, 8, 16	–	10	34.0
	FIVES	$10^{-6}$ – $10^{-1}$	$10^{-6}$ – $10^{-1}$	2, 4, 8, 16, 32, 64	first width 8, 16, 32	1	108.6
	VessMap	$10^{-5}$ – $10^{-2}$	$10^{-5}$ – $10^{-1}$	2, 4, 8, 16, 32	first width 8, 16, 32	1	1.2
ConvUNeXt	DRIVE	$10^{-6}$ – $10^{-2}$	$10^{-6}$ – $10^{-1}$	2, 4, 8	base width 16, 32, 64	1	13.4
	HRF	$10^{-6}$ – $10^{-2}$	$10^{-6}$ – $10^{-1}$	2, 4, 8, 16	–	10	25.1

Appendix

Table A.1: Hyperparameter search. Learning rate and weight decay: log-uniform in the ranges shown. CNN architecture parameter: LW-Net , width of the first of its three levels (doubled at each level); ConvUNeXt , width of the first stage; ConvNeXt , stochastic-depth (drop-path) rate. Val.: validation interval in epochs. GPU h: sum of the trial durations.

Model	Version	Input	Patch	Tokens	Emb.	Depth	Heads	MLP	Window	Params	Dice	$R_{\mathrm{thin}}$	$P_{\mathrm{thin}}$
MedSAM	original	1024	16	$64\times 64$	768	12	12	4	14	93.7	Table B.1
	compact, DRIVE	256	4	$64\times 64$	256	4	8	2	14	7.9	$77.7{\pm}2.5$	$43.4{\pm}13.5$	$69.3{\pm}8.3$
	compact, HRF	256	4	$64\times 64$	128	6	4	2	14	6.0	$78.3{\pm}5.9$	$23.7{\pm}15.3$	$43.3{\pm}11.6$
	compact, FIVES	256	8	$32\times 32$	256	8	8	6	7	13.5	$87.8{\pm}9.0$	$32.0{\pm}11.3$	$41.9{\pm}10.7$
	compact, VessMap	256	4	$64\times 64$	384	6	4	6	7	20.6	$91.7{\pm}2.8$	$77.5{\pm}18.2$	$74.5{\pm}14.5$
LiteMedSAM	original	256	–	–	64,128,160,320	2,2,6,2	2,4,5,10	4	7,7,14,7	9.8	Table B.1

Appendix

Table A.2: Selected compact models and, for comparison, the original MedSAM and LiteMedSAM . Input and Patch in px. Tokens: patch grid of one window, calculated as image size divided by patch size. Emb.: embedding dimension. MLP: ratio of the hidden width of the MLP block. Window: size of the attention window in patches. Params: number of parameters in millions.

Patch (px)	Tr.	Dice	Tr.	Dice	Tr.	Dice	Tr.	Dice
	DRIVE		HRF		FIVES		VessMap
1	0	OOM	0	OOM	0	OOM	0	OOM
2	0	OOM	0	OOM	0	OOM	0	OOM
4	18	76.8	44	78.9	4	85.3	37	92.1
8	16	70.2	13	75.0	69	89.6	32	90.5
16	56	15.7	16	14.0	8	82.7	6	87.0

Appendix

Table A.3: Patch sizes in the search of compact MedSAM . Number of trials that reached a validation epoch (Tr.) and their best validation Dice (%). Bold indicates the best performance for each dataset, which set the patch sizes used in the experiments. OOM: every trial ran out of GPU memory.

Dataset	Metric	FT	Scratch	$\Delta$	FT	Scratch	$\Delta$	CNN
		LiteMedSAM			MedSAM
DRIVE ( $n=20$ )	Dice	$82.3\pm 1.9$	$81.4\pm 2.4$	$+0.9^{*}$	$81.3\pm 1.3$	$78.7\pm 1.6$	$+2.6^{*}$	$81.1\pm 2.4$ L
	$R_{\mathrm{thin}}$	$64.3\pm 11.6$	$59.9\pm 13.4$	$+4.4^{*}$	$56.9\pm 11.8$	$39.3\pm 11.2$	$+17.6^{*}$	$57.5\pm 12.1$ L
	$P_{\mathrm{thin}}$	$74.5\pm 5.2$	$76.2\pm 5.8$	$-1.7$	$73.0\pm 6.1$	$76.2\pm 7.0$	$-3.2^{*}$	$80.8\pm 8.6$ U
HRF ( $n=14$ )	Dice	$80.6\pm 5.9$	$79.5\pm 5.1$	$+1.1$	$81.7\pm 4.5$	$79.7\pm 5.4$	$+2.0^{*}$	$79.7\pm 5.1$ U
	$R_{\mathrm{thin}}$	$40.1\pm 15.7$	$30.2\pm 13.6$	$+9.9^{*}$	$26.9\pm 17.2$	$28.0\pm 16.1$	$-1.1$	$25.4\pm 13.5$ N

Appendix

Table B.1: Effect of pre-training. LiteMedSAM and MedSAM fine-tuned (FT) and trained from scratch, and the best CNN. L: LW-Net , U: ConvUNeXt , N: ConvNeXt .

\Delta

shows the FT value minus Scratch, with an asterisk when the difference is significant. Bold indicates the best value, including the CNN. Underlined values are not significantly different from the best of the same row.

	DRIVE			HRF			FIVES			VessMap
Pair	Dice	$R_{\mathrm{thin}}$	$P_{\mathrm{thin}}$	Dice	$R_{\mathrm{thin}}$	$P_{\mathrm{thin}}$	Dice	$R_{\mathrm{thin}}$	$P_{\mathrm{thin}}$	Dice	$R_{\mathrm{thin}}$	$P_{\mathrm{thin}}$
LW vs CU	0.58	0.054	0.19	0.004 -	1.00	0.54	$<$ 0.001 -	$<$ 0.001 -	$<$ 0.001 +	1.00	1.00	1.00
LW vs CN	$<$ 0.001 +	$<$ 0.001 +	0.010 +	1.00	1.00	0.59	$<$ 0.001 -	$<$ 0.001 -	0.059	$<$ 0.001 +	$<$ 0.001 +	0.75
LW vs LM	0.005 -	$<$ 0.001 -	$<$ 0.001 +	0.005 -	0.004 -	0.25	$<$ 0.001 -	$<$ 0.001 -	$<$ 0.001 +	1.00	0.076	1.00
LW vs LM-S	0.44	0.008 -	0.025 +	0.004 -	0.62	1.00	$<$ 0.001 -	0.33	1.00	$<$ 0.001 +	0.002 +	0.15
LW vs LM-C	0.77	$<$ 0.001 -	$<$ 0.001 +	0.005 -	0.47	1.00	$<$ 0.001 -	$<$ 0.001 -	$<$ 0.001 +	0.011 +	1.00	0.073

Appendix

Table C.1: Holm-adjusted p-values of the pairwise tests. A

+

superscript indicates that the first model of the pair is significantly better (

p<0.05

). A

-

is used when the second configuration is better. LW: LW-Net ; CU: ConvUNeXt ; CN: ConvNeXt ; LM, LM-S, LM-C: LiteMedSAM fine-tuned, from scratch and compact; M, M-S, M-C: MedSAM fine-tuned, from scratch and compact.

	DRIVE		HRF		FIVES		VessMap
Pair	$R_{\mathrm{cl}}$	$P_{\mathrm{cl}}$	$R_{\mathrm{cl}}$	$P_{\mathrm{cl}}$	$R_{\mathrm{cl}}$	$P_{\mathrm{cl}}$	$R_{\mathrm{cl}}$	$P_{\mathrm{cl}}$
LW vs CU	0.048 +	0.36	1.00	0.23	$<$ 0.001 -	$<$ 0.001 -	0.025 +	0.29
LW vs CN	$<$ 0.001 +	1.00	1.00	0.026 +	$<$ 0.001 -	$<$ 0.001 -	$<$ 0.001 +	1.00
LW vs LM	$<$ 0.001 -	$<$ 0.001 +	0.001 -	1.00	$<$ 0.001 -	$<$ 0.001 -	0.002 -	0.084
LW vs M	0.70	$<$ 0.001 +	0.043 -	0.27	$<$ 0.001 -	$<$ 0.001 -	0.037 -	1.00
CU vs CN	0.003 +	0.43	1.00	0.001 +	$<$ 0.001 -	0.14	0.028 +	0.55

Appendix

Table C.2: Holm-adjusted p-values for Table 2 (top) and Table 3 (bottom). LW: LW-Net ; CU: ConvUNeXt ; CN: ConvNeXt ; LM: LiteMedSAM fine-tuned; M: MedSAM fine-tuned. See the caption of Table C.1 for the meaning of the

+

and

-

superscripts.

Retinal Vessel Segmentation

Retinal vessel segmentation identifies blood vessels in retinal images to support vascular measurement, disease analysis, and screening.

See all 25 papers

HMS-VesselNet: Hierarchical Multi-Scale Attention Network with Topology-Preserving Loss for Retinal Vessel Segmentation
Mar 23, 2026Amarnath RRetinal Vessel Segmentation Multi-Scale Feature Fusion
United Institute of Technology, Coimbatore, India
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
U-PEN Mamba: Progressive Expansion with Selective State-Space Modeling for Efficient Retinal Vessel Segmentation
Sep 21, 2026Abel A. Reyes-Angulo, Sidike Paheding, Vijayan K. Asari +2Mamba Retinal Vessel Segmentation
Michigan Technological University, Houghton, MI, USA · Fairfield University, Fairfield, CT, USA · University of Dayton, Dayton, OH, USA +1

Mar 23, 2026J/K move · Enter open · S save

HMS-VesselNet: Hierarchical Multi-Scale Attention Network with Topology-Preserving Loss for Retinal Vessel Segmentation

Amarnath R

United Institute of Technology, Coimbatore, India

Retinal Vessel Segmentation Multi-Scale Feature Fusion Topology-Preserving Image Segmentation

Open paper AskPDF arXiv

Retinal vessel segmentation methods based on standard overlap losses tend to miss thin peripheral vessels because these structures occupy very few pixels and have low contrast against the background. We propose HMS-VesselNet, a hierarchical multi-scale network that processes fundus images across four parallel branches at different resolutions and combines their outputs using learned fusion weights. The training loss combines Dice, binary cross-entropy, and centerline Dice to jointly optimize area overlap and vessel continuity. Hard example mining is applied from epoch 20 onward to concentrate gradient updates on the most difficult training images. Tested on 68 images from DRIVE, STARE, and CHASE_DB1 using 5-fold cross-validation, the model achieves a mean Dice of 88.72 +/- 0.67%, Sensitivity of 90.78 +/- 1.42%, and AUC of 98.25 +/- 0.21%. In leave-one-dataset-out experiments, AUC remains above 95% on each unseen dataset. The largest improvement is in the recall of thin peripheral vessels, which are the structures most frequently missed by standard methods and most critical for early detection of diabetic retinopathy.

Abstract

Retinal Vessel Segmentation

Explore similar work

HMS-VesselNet: Hierarchical Multi-Scale Attention Network with Topology-Preserving Loss for Retinal Vessel Segmentation

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

U-PEN Mamba: Progressive Expansion with Selective State-Space Modeling for Efficient Retinal Vessel Segmentation