REVIEW 4 major objections 5 minor 5 references
A Comprehensive Review of U-Net and Its Variants: Advances and Applications in Medical Image Segmentation
T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash
Pith's one-line read A review of U-Net and its variants claims that four mechanisms — jump connections, residual connections, 3D U-Net, and transformers — organize medical image segmentation, and maps each dataset modality to the models best suited to it.
desk verdict Broad but sloppy U-Net survey: a useful starting map for newcomers, but the table inconsistencies and garbled equations make it unreliable as a reference until major corrections. 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 device is the four-mechanism taxonomy, with each mechanism named and defined by the paper. Jump connections are the skip connections that shuttle high-resolution features from encoder to decoder. Residual connections are identity shortcuts that let gradients flow and depth grow without degradation. The 3D U-Net is the volumetric form of the architecture, using 3D convolutions to exploit spatial structure. The transformer mechanism replaces or augments convolutions with self-attention to capture global dependencies. The review uses this taxonomy as the spine for its comparison tables (Tables 4 to 6), its dataset-to-model mapping (Table 1), and its strengths-and-weaknesses summary (Table 7).
What would settle it
Look up the original papers cited in Tables 4 to 6 and check whether the listed Dice or Jaccard values match what those papers report under the same datasets and evaluation protocols. A concrete check: UNETR appears with year 2021 and Dice values 89.1 and 96.4 in Table 5, and with year 2022 and Dice values 87.2, 83.28, and 92.83 in Table 6; reconciling these two entries would test whether the review's comparison tables are reliable.
Extended reading notes
Core claim
On the paper's own terms, the contribution is an organizing claim about the literature: the U-Net's dominance in medical image segmentation is best explained by four structural enhancement mechanisms. The jump-connection mechanism preserves fine detail by carrying encoder features to the decoder; the residual-connection mechanism, borrowed from ResNet and dense networks, prevents degradation and improves feature reuse; the 3D U-Net mechanism extends the architecture to volumetric data; and the transformer mechanism adds self-attention for global context, typically in hybrid U-Net/Transformer designs such as TransUNet, Swin-UNet, and UNETR. The review maps these mechanisms to imaging modalities and datasets, compares reported performance, and condenses the field's open problems into four limitations and four improvement directions. Its claim is that this four-part taxonomy captures the design space of U-Net variants and points to future work.
Load-bearing premise
The review assumes that the numeric results copied into Tables 4 to 6 are accurate, comparable, and correctly attributed to the cited papers; if those numbers are wrong or mislabeled, the advantages, disadvantages, and improvement directions lose their evidentiary basis.
Editorial extensions
If this is right
- A newcomer to medical image segmentation can use the modality-to-model map as a first-choice guide for picking a U-Net-style baseline.
- New U-Net variants can be classified by which of the four mechanisms they add, making the literature easier to compare.
- The four mechanisms are combinable; the review's own examples show residual, 3D, and transformer ideas appearing together in one network.
- Transformer-based variants are presented as the main answer to U-Net's limited global context, but at the cost of higher computational demand.
- The review's four improvement directions—feature extraction, receptive field, spatial information, and computational cost—give a checklist for designing the next variant.
Reading between the lines
- If the taxonomy is accepted, the right benchmark question for a new variant is not 'does it beat U-Net?' but 'which mechanism does it add, and what does that mechanism cost?'
- The paper leaves implicit that a fair comparison of the variants in Tables 4 to 6 would require identical datasets, preprocessing, and metric definitions; those are not guaranteed across the cited papers.
- The dataset-to-model mapping could be extended into a decision rule that also considers annotation budget and image resolution, but the review does not attempt that.
- The prominence of transformers in the review suggests that the next synthesis step—combining transformers with 3D and residual mechanisms in a single efficient network—is already visible in the literature.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This manuscript is a narrative review of U-Net and its variants for medical image segmentation. It opens by classifying medical image datasets by modality and briefly describing U-Net's architecture. It then surveys several families of U-Net modifications: networks with multiple encoders, multiple or cascaded U-Nets, 3D U-Nets, and a section on semi-supervised and unsupervised domain adaptation. The paper proposes that U-Net improvements can be distilled into four core mechanisms: skip connections, residual connections, 3D processing, and transformers. It tabulates reported performance of selected models in Tables 4–6, discusses challenges and solutions, and closes with future research directions including interpretability and the Segment Anything Model.
Significance. The paper covers a wide range of U-Net variants and recent transformer-based models, and a reliable synthesis would be a useful reference for researchers entering the field. The list of challenges and improvement strategies in Section 5 and Table 7 provides a compact mental model. However, the value of the review depends on the accuracy and comparability of the numbers in Tables 4–6 and on a reproducible selection methodology. At present, internal inconsistencies in those tables and the lack of a stated search or inclusion process substantially weaken the paper's status as a 'systematic summary.'
major comments (4)
- [Tables 5 and 6] The UNETR entries are inconsistent across the two tables and within Table 5 itself. Table 5 lists UNETR twice, once with year 2021 and once with year 2022, giving identical Dice values of 89.1% and 96.4% on 'BCTV' and MSD. Table 6 lists UNETR, year 2022, with DSC values of 87.2%, 83.28%, and 92.83% on Synapse, BTCV, and ACDC. These discrepancies in year, dataset names, and scores are not addressed. Because the comparative claims in Section 5 and the 'Advantages'/'Disadvantages' columns are built on these transcriptions, the numerical evidence base is unreliable.
- [Section 4] The metric definitions in Section 4 are garbled. The DSC formula appears as '2 pre tru pre tru Area Area DSC Area Area = +', the JI formula splits across lines incorrectly, the IoU expression is written as 'rp rp Area AreaIoU Area Area=', and the Hausdorff distance formula is an unreadable string. The notation such as 'preArea', 'truArea', 'Arear', and 'Areap' is not defined. Since Tables 4–6 report these metrics, correct and legible definitions are essential for the paper's reference function.
- [Systematic methodology] The abstract and title claim a 'systematic summary' and a 'comprehensive review,' but the manuscript states no search strategy, inclusion/exclusion criteria, or quality assessment for the surveyed literature. The selection of the specific models in Tables 4–6 appears arbitrary, and the reader cannot determine which studies were considered. Without a reproducible methodology, the four-mechanism synthesis in Table 7 cannot be verified as a systematic result rather than an ad hoc categorization.
- [Section 3.2 and Table 7] There is a structural inconsistency between the stated four core mechanisms and the paper's own organization. The abstract and Table 7 identify skip connections, residual connections, 3D U-Net, and transformers as the four central mechanisms, but Section 3.2 is arranged as 'Residual Neural Network Mechanism,' 'Dense Convolutional Network Mechanism,' 'Transformer mechanism,' and 'Attention mechanism,' with 3D U-Net treated as a model family in Section 2.3 rather than as an improvement mechanism. Dense connections and attention are largely omitted from Table 7. This mismatch makes the central synthesis difficult to follow.
minor comments (5)
- [Section 1] The citation 'J et al.' for the FCN paper should be 'Long et al.'; reference [1] itself is correct but the in-text attribution is wrong.
- [Table 1] The entry 'Montgomery and Shenzhen []Chest X-ray Sets' contains an empty reference placeholder; please complete the citation.
- [Section 5] In the suggestion paragraph, 'UNETR++ [32]' is a citation error: reference [32] is MultiResUNet, not UNETR++. The same paragraph also refers to 'UNETR++[81]' and 'UNETR++ [81]' with inconsistent formatting; verify the intended reference.
- [Throughout] There are numerous typographical errors and inconsistent names, for example 'Dense-Inception U-Ne[74]t', 'Cascaded Bionet[36]' for BiO-Net, and 'UNet-Sharp' without a reference. A careful proofreading pass is needed.
- [Section 4] The notation in Section 4 uses 'Area_pre' and 'Area_tru' without defining them as the predicted and ground-truth regions, and the confusion matrix in Figure 26 has a typo in 'F_N' definition; consider a cleaner presentation.
Circularity Check
No circularity: the review categorizes external results; the UNETR table inconsistencies are accuracy issues, not derivation-by-construction.
full rationale
This is a survey/review paper, not a derivation. Its central claims are classifications and summaries of U-Net variants (jump connection, residual connection, 3D U-Net, transformer mechanism), supported by descriptions of external published architectures. There is no internal derivation chain in which a claimed output is defined in terms of an input, no parameter fitted to a subset and then renamed as a prediction, and no load-bearing self-citation chain: the references cite prior work by other authors (Ronneberger, Cicek, Milletari, etc.), and the present authors do not invoke their own prior theorem as a premise. The numerical evidence in Tables 4-6 is transcribed from other papers, and indeed contains inconsistencies (e.g., UNETR appears with conflicting years and Dice values in Tables 5 and 6), but this is a correctness/verifiability concern, not circularity. The paper's own limitation statement in Section 5.3 ('The proposed solution in this study combined with the study of the problem of scarce data for medical image annotation does not completely eliminate this fundamental drawback') is an admitted scope limitation, not a circular step. The recommendations in Section 5 are advisory syntheses of external results, not predictions forced by construction. Accordingly, no circular step is present, and the review is self-contained as a descriptive survey.
Assumptions & free parameters
assumptions (1)
- domain assumption Reported quantitative results in the surveyed papers are trustworthy and comparable across studies
Cite this review
Pith. "Pith review of A Comprehensive Review of U-Net and Its Variants: Advances and Applications in Medical Image Segmentation." pith.science (2026). https://pith.science/paper/IGSMLBTU
@misc{pith2026250206895,
author = {Pith},
title = {Pith review of: A Comprehensive Review of U-Net and Its Variants: Advances and Applications in Medical Image Segmentation},
year = {2026},
howpublished = {\url{https://pith.science/paper/IGSMLBTU}},
note = {Machine review of arXiv:2502.06895}
}
read the original abstract
Medical images often exhibit low and blurred contrast between lesions and surrounding tissues, with considerable variation in lesion edges and shapes even within the same disease, leading to significant challenges in segmentation. Therefore, precise segmentation of lesions has become an essential prerequisite for patient condition assessment and formulation of treatment plans. Significant achievements have been made in research related to the U-Net model in recent years. It improves segmentation performance and is extensively applied in the semantic segmentation of medical images to offer technical support for consistent quantitative lesion analysis methods. First, this paper classifies medical image datasets on the basis of their imaging modalities and then examines U-Net and its various improvement models from the perspective of structural modifications. The research objectives, innovative designs, and limitations of each approach are discussed in detail. Second, we summarize the four central improvement mechanisms of the U-Net and U-Net variant algorithms: the jump-connection mechanism, residual-connection mechanism, 3D-UNet, and transformer mechanism. Finally, we examine the relationships among the four core enhancement mechanisms and commonly utilized medical datasets and propose potential avenues and strategies for future advancements. This paper provides a systematic summary and reference for researchers in related fields, and we look forward to designing more efficient and stable medical image segmentation network models based on the U-Net network.
Reference graph
Works this paper leans on
-
[3]
Improvement methods and strategies To improve the original U -Net model in terms of network performance, efficiency, accuracy, and generalizability, several enhancements have been introduced. These improvements encompass various aspects of the U -Net architecture, including data augmentation, convolution operations, downsampling operations and upsampling ...
work page 2023
-
[4]
Future Directions. In recent years, U-Net and variant network models have shown excellent performance in medical dataset segmentation. With advancements in GPU computing and their increasing applications, segmentation models and transformers have been combined, and the model's segmentation accuracy and overall performance have greatly improved. This study...
-
[5]
Additionally, manual annotation has difficulty avoiding subjective errors
Challenges and Solutions 5.1 Limitations of medical datasets Most segmentations for medical datasets rely on supervised deep neural network models, and high -quality medical image segmentation annotation requires expert -level knowledge, which demands a significant amount of time and incurs high costs. Additionally, manual annotation has difficulty avoidi...
work page 2020
-
[6]
W-Net (Xia and Kulis,2017) Multiple parallels combine U-shaped networks in a parallel manner, assign different functions to different U-Net networks, and input different types of images. Combining U-Net with other models. Zhao et al. proposed a deep learning model Triple U-net [22]. specifically designed for the segmentation of pathological cancer nuclei ...
work page 2017
-
[8]
References: 1 J, L., E, S., T, D.: ‘Fully convolutional networks for semantic segmentation’IEEE, 2015. 2 Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: ‘3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation’, in ‘Medical Image Computing and Computer- Assisted Intervention – MICCAI 2016’ (Springer International Publ...
arXiv 2018
Reviewed August 8, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.