REVIEW 4 major objections 5 minor 1 references
Dual encoding feature filtering generalized attention UNET for retinal vessel segmentation
T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read A dual-encoder U-Net that adds an enhanced high-frequency input stream claims better retinal vessel segmentation and stronger cross-dataset generalization than single-encoder baselines.
desk verdict A plausible engineering combination, but the headline generalization claim is compromised by an unspecified augmentation reference set and contradicted by the paper's own tables. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the domain-invariant input stream. Each raw image $I$ is converted to a high-frequency map by subtracting a local average, $H(x,y)=I(x,y)-L_{\mathrm{avg}}(x,y)$, then amplifying it by a global factor $G$ computed from the variance of $H$. Encoder-2 consumes $I_{\mathrm{enhanced}}=G\cdot H$, while Encoder-1 consumes the raw image. Before the bottleneck, the Feature Filtering Fusion (FFF) module applies channel attention to the domain-specific features ("what to emphasize") and spatial attention to the domain-invariant features ("where to look"), then convolves the concatenation. Replacing each skip connection, the Feature Reconstructing Fusion (FRF) module concatenates low-level features from both encoders, splits them into dilated-context and local paths, and uses fused high-level features to produce attention weights for the reconstructed low-level map. JESB balances training sets by clustering binary vessel masks with Jaccard distance and synthesizing samples for small clusters, while SOTA-CSA augments by mixing each image with the color statistics of reference datasets.
What would settle it
A direct check is to compare the pixel-value distributions of the enhanced high-frequency maps at vessel locations across DRIVE and STARE: if those distributions are no closer to each other than the raw color distributions are, the premise that these components are domain-invariant fails.
Extended reading notes
Core claim
The paper's discovery claim is that the high-frequency part of a fundus image is domain-invariant, so a second encoder fed an enhanced version of that signal can give a U-Net both richer vessel features and substantially better generalization. DEFFA-Unet is this dual-encoder design: Encoder-1 reads the raw image, Encoder-2 reads $I_{\mathrm{enhanced}}=G\cdot H$, where $H$ is the image minus its local average and $G$ is a variance-derived enhancement factor. A Feature Filtering Fusion (FFF) module applies channel attention to the domain-specific stream and spatial attention to the domain-invariant stream before the bottleneck, and a Feature Reconstructing Fusion (FRF) module replaces skip connections with attention-weighted reconstruction from both low- and high-level features. With Jaccard-distance-based synthetic balancing and color-statistics augmentation, the paper reports top or near-top accuracy, recall, specificity, precision, Dice, IoU, AUC and MCC on DRIVE, CHASEDB1, STARE, HRF and IOSTAR, and reports that leave-one-out cross-domain and cross-modality tests beat or match multi-domain methods without target-domain fine-tuning.
Load-bearing premise
The load-bearing premise is that the high-frequency components of a fundus image carry the vessel structure and stay roughly constant when the acquisition device changes, and the paper asserts this without directly measuring that stability.
Editorial extensions
If this is right
- A vessel segmenter trained on one fundus dataset could be deployed on unseen datasets without target-domain fine-tuning, reducing the need for new annotations for every scanner or patient population.
- The attention-guided replacement of skip connections is designed to cut false positives, the failure mode the paper says is most costly in clinical screening.
- The JESB balancing and SOTA-CSA augmentation recipes use only cheap image-level operations and could be dropped into other segmentation pipelines.
- The model matches or exceeds multi-domain methods while using a small parameter footprint (2.85 MB in the reported tables), so the reported generalization gain is not bought by model scale.
Reading between the lines
- Beyond the paper: if high-frequency domain invariance is a general property of curvilinear structures, the same second-stream design should transfer to other medical targets such as corneal nerves or coronary vessels; this is untested here.
- Beyond the paper: the reported gains bundle the dual encoder, both fusion modules, and the augmentation schemes together; isolating each contribution with a noise-fed second encoder would show which ingredient actually carries the generalization.
- Beyond the paper: because SOTA-CSA only needs color statistics from a reference set, using unlabeled statistics from the target device would turn the method into a lightweight unsupervised adaptation step.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes DEFFA-Unet, a dual-encoder U-Net variant for retinal vessel segmentation, together with two data-processing methods: JESB (Jaccard-based synthetic balancing) and SOTA-CSA (color-statistics-based augmentation), and a high-frequency domain-invariant input branch. The authors report experiments on DRIVE, STARE, CHASE, HRF, and IOSTAR, claiming improved segmentation accuracy and, especially, cross-domain generalization. The paper contains model descriptions, quantitative comparisons with prior methods, cross-domain evaluations, and an ablation study.
Significance. If the claims were substantiated, a lightweight dual-encoder network with domain-invariant preprocessing and targeted augmentation would constitute a practically useful contribution to retinal vessel segmentation, and the cross-domain generalization results would be of interest to the medical imaging community. The paper does provide a well-structured architecture description and a reasonably broad experimental setup. However, the reported evidence is internally inconsistent and does not support the stated conclusions; the proposed method is not numerically superior to several compared methods on the primary metric in the paper's own tables, and the cross-domain augmentation source is ambiguous. The potential significance is therefore not realized in the current manuscript.
major comments (4)
- [Section 3.3.1, Tables 3 and 4] The claim that 'our proposed method achieved the best scores in all reported metrics' is directly contradicted by the paper's own tables: on DRIVE the proposed method's DSC is 0.8247 versus 0.8310 for CSGNet and 0.8303 for DPFNet; on STARE the proposed DSC is 0.8111 versus 0.8516 for CSUNet and 0.8493 for CSGNet; on CHASE the proposed DSC is 0.8090 versus 0.8302 for DPFNet; on IOSTAR the proposed DSC is 0.8182 versus 0.8516 for SCS-Net. The proposed method also does not achieve the best recall in several rows. The manuscript provides no error bars, confidence intervals, or significance tests for any of these comparisons, so the abstract's and Section 1's claims of 'significantly outperforms' are not supported by the presented evidence.
- [Section 2.1.2, Equations (4)-(6)] The 'SOTA dataset' D used to compute the color statistics μ_C and σ_C is never defined. The cross-domain experiments in Tables 5 and 6 train on one dataset (e.g., DRIVE or IOSTAR) and test on others. If the 'SOTA dataset' includes or overlaps with the target test sets, then SOTA-CSA injects target-domain color statistics into the training data, and the reported cross-domain gains are a form of leakage rather than evidence of generalization. The paper must unambiguously state that the statistics are computed from the training source domain only, and preferably verify this by ablating the augmentation source. Without this clarification, the central cross-domain generalization claim is uninterpretable.
- [Section 2.1.3, Equations (7)-(10)] The core assumption underlying the second encoder, that high-frequency components of retinal images are domain-invariant and preserve vessel structure across datasets, is asserted without proof or empirical validation. The entire generalization benefit is attributed to this branch, and the 'w/o D-I' ablation in Table 6 lacks error bars or repeated runs. The authors should provide evidence for this assumption, for instance by analyzing the distribution of high-frequency features across domains or by comparing the proposed high-frequency preprocessing to the original FDA-based approach with a controlled experiment.
- [Section 4 (Discussion)] The Discussion states: 'It is crucial to clarify that this paper does not claim superiority of the proposed method over others.' This directly contradicts the Abstract ('the proposed method significantly outperforms the compared methods in cross-validation model generalization'), the Introduction, and Section 3.3.1, which claim the best scores in all reported metrics. This is not a minor wording issue: it makes the paper's central claim incoherent, and it also suggests that the authors themselves do not stand behind the headline results. The manuscript needs to be rewritten so that the claims match the actual numerical evidence.
minor comments (5)
- [Table 1] There is a typo 'CHSE' in Section 2.1.1, presumably meaning CHASE; also the dataset overview in Table 1 lists CHASE as having 28 total images with 'R:28, H:0', but the text in Section 3.1 says CHASEDB1 was divided into 20 training and 8 testing images, which is inconsistent with the total count of 28 and the healthy/retinopathy split.
- [Table 5] The header 'Trained on DRIVE (D) Leave-1 (D, S, C, H)' is confusing and appears to mix two separate training protocols; this makes it difficult to interpret the baseline comparisons and should be clarified.
- [References] In Table 5, the method labeled 'HGC-Net' is cited as [45], but the reference list associates [45] with SCS-Net and [46] with the self-supervision boosted method; the citation numbering appears inconsistent.
- [Table 5] The MCC value '0.68774' is reported with an extra digit compared to the other entries; this looks like a transcription error.
- [Table 6] The abbreviation 'w/o D-I' is not defined in the table caption; the text later explains it as 'without domain-invariant feature guidance', but the caption should state this.
Circularity Check
SOTA-CSA's reference 'SOTA dataset' is unspecified; if it includes the target test sets, the Tables 5-6 cross-domain gains are fitted to target color statistics rather than predicted.
-
fitted input called prediction
[Section 2.1.2 (Eqs. 4-6); Section 3.3.2 (Tables 5-6)]
"By utilizing SOTA datasets as reference, proposed method ensure both data consistency and data diversity which enhance model generalization on unseen data. ... The augmentation is applied based on calculated SOTA color statistics on the source dataset, S in Equation (6). ... our method achieved higher AUC and MCC in all three datasets without any fine-tuning with target datasets."
Tables 5-6 train on one dataset (DRIVE or IOSTAR) and test on STARE/CHASE/HRF/DRIVE. SOTA-CSA fits μ_C, σ_C from a set designated only as 'SOTA datasets' and applies p_aug = α((p−μ_C)/σ_C)+(1−α)p to source training images. The paper never states whether that reference set is restricted to the source training domain or includes the target test datasets. If it includes the target, target color statistics are injected into training, so the large CSA-attributed cross-domain margin (STARE AUC 0.9415 without vs 0.9834 with CSA in Table 6) is fitted to the target distribution, not predicted. The phrase 'without any fine-tuning with target datasets' rules out fine-tuning but not the use of target statistics in augmentation; no ablation or code resolves the ambiguity.
full rationale
This is an empirical architecture paper, so most components cannot be circular in a derivation sense: the dual-encoder, FFF/FRF attention modules, JESB balancing, and high-frequency preprocessing are defined independently of the test labels, and the high-frequency domain-invariance claim is an unproven assumption rather than a circular step. There is no self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in from the authors' own prior work. The single substantive circularity risk is SOTA-CSA: Eqs. (4)-(6) compute statistics from an unspecified 'SOTA dataset' and apply them to the source set, while the paper's headline cross-domain results (Tables 5-6) are evaluated on datasets that could be included in that reference set. Because the text never states whether the target/test sets contribute statistics, the central generalization claim is not fully self-contained: if the target is included, the reported gains are leakage rather than generalization. The ambiguity prevents a definitive score of 6+; the Discussion's disclaimer that 'this paper does not claim superiority' is an internal consistency issue, not circularity. I therefore assign a moderate score of 4.
Assumptions & free parameters
free parameters (4)
- alpha_loss =
0.25
- alpha_enhance =
unspecified
- window_size =
unspecified
- alpha_blend =
random in [0.7, 1.0]
assumptions (3)
- domain assumption High-frequency components of fundus images are domain-invariant and encode vessel structure across different acquisition conditions.
- domain assumption The FOV masks generated by W-Net for CHASE and STARE are accurate substitutes for the unavailable official masks.
- domain assumption KMeans clustering with silhouette score on Jaccard distances yields a valid balancing partition for synthetic data generation.
Cite this review
Pith. "Pith review of Dual encoding feature filtering generalized attention UNET for retinal vessel segmentation." pith.science (2026). https://pith.science/paper/QTGZI2SJ
@misc{pith2026250602312,
author = {Pith},
title = {Pith review of: Dual encoding feature filtering generalized attention UNET for retinal vessel segmentation},
year = {2026},
howpublished = {\url{https://pith.science/paper/QTGZI2SJ}},
note = {Machine review of arXiv:2506.02312}
}
read the original abstract
Retinal blood vessel segmentation is crucial for diagnosing ocular and cardiovascular diseases. Although the introduction of U-Net in 2015 by Olaf Ronneberger significantly advanced this field, yet issues like limited training data, imbalance data distribution, and inadequate feature extraction persist, hindering both the segmentation performance and optimal model generalization. Addressing these critical issues, the DEFFA-Unet is proposed featuring an additional encoder to process domain-invariant pre-processed inputs, thereby improving both richer feature encoding and enhanced model generalization. A feature filtering fusion module is developed to ensure the precise feature filtering and robust hybrid feature fusion. In response to the task-specific need for higher precision where false positives are very costly, traditional skip connections are replaced with the attention-guided feature reconstructing fusion module. Additionally, innovative data augmentation and balancing methods are proposed to counter data scarcity and distribution imbalance, further boosting the robustness and generalization of the model. With a comprehensive suite of evaluation metrics, extensive validations on four benchmark datasets (DRIVE, CHASEDB1, STARE, and HRF) and an SLO dataset (IOSTAR), demonstrate the proposed method's superiority over both baseline and state-of-the-art models. Particularly the proposed method significantly outperforms the compared methods in cross-validation model generalization.
Reference graph
Works this paper leans on
-
[1]
2025 年 1 月 第 62 卷 第 1 期 Jan . 2025 Vol . 62 No . 1 四川大学学报 （ 自然科学版 ） Journal of Sichuan University （ Natural Science Edition ） 双编码特征过滤泛化注意力 UNET 视网膜血管分割 陶西洛 1 ， 吴达文 2 ， 唐青青 2 ， 赵凯洋 3 ， 尹 腾 1 ， 李彦霏 1 ， 尚文一 1 ， 刘晶玉 1 ， 张海仙 1 （ 1 . 四川大学计算机学院机器智能实验室 ， 成都 610064 ； 2 . 四川大学华西医院眼科 ， 成都 610041 ； 3 . 四川大学华西临床医学院 ， 成都 610041 ） 摘 要 : 视 网 膜 血 管 的 形 态 与 功 能 改 变 对 临 床 中...
work page 2025
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.