Pith. sign in

REVIEW 4 major objections 7 minor 24 references

Convolutional Neural Network Segmentation for Satellite Imagery Data to Identify Landforms Using U-Net Architecture

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a U-Net with dropout and the Adam optimizer can segment satellite landform images, reporting 69.62% Dice and 90.53% model accuracy on 5,000 images.

desk verdict Standard U-Net on landforms, but the reported metrics are not shown to be held-out and the comparison table contradicts the outperformance claim. read the letter →

arxiv 2502.05476 v1 pith:QMX6UXWU submitted 2025-02-08 cs.CV

Mitul Goswami , Sainath Dey , Aniruddha Mukherjee , Suneeta Mohanty , Prasant Kumar Pattnaik This is my paper · ORCID
classification cs.CV
keywords U-NetConvolutionalNeuralNetworkSemanticSegmentationSatelliteImageryLandformidentificationDicecoefficientDropoutAdamoptimizer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper attempts to establish that a U-Net convolutional neural network, equipped with dropout regularization and the Adam optimizer, can take preprocessed satellite landform images and produce pixel-level segmentation maps that distinguish forests, deserts, mountains, and water. On 5,000 preprocessed satellite landform images, the model is reported to reach a Dice coefficient of 69.62% and a Model Accuracy of 90.53%. The paper presents these numbers as evidence that the architecture is a practical, adaptable tool for landform recognition in applications such as autonomous driving, disaster management, and land-use planning. The central claim is that a standard encoder-decoder U-Net, trained with binary cross-entropy and evaluated with the Dice coefficient, is a workable recipe for landform segmentation.

What carries the argument

The carrying mechanism is the U-Net architecture itself: a contracting path of Conv2dBlocks (two convolutions, batch normalization, and ReLU) with max-pooling extracts hierarchical features while downsampling, and an expanding path of Conv2DTranspose layers upsamples those features back to the input resolution. Skip connections concatenate corresponding encoder feature maps with decoder feature maps, preserving fine spatial detail for pixel-level classification. Dropout at rate 0.07 regularizes the network, the Adam optimizer adapts per-parameter learning rates, and a final sigmoid layer with binary cross-entropy loss produces a per-pixel probability map. The paper's quantitative measure is the Sørensen-Dice coefficient, defined as $2TP/(2TP+FN+FP)$, which it uses to score overlap between predicted and ground-truth masks.

What would settle it

Re-run the described U-Net (dropout rate 0.07, Adam optimizer, binary cross-entropy loss) on a fixed train/test split of the same 5,000-image landform dataset and compute the Dice coefficient on the held-out test images; if the test Dice falls substantially below 69.62%, the reported segmentation quality would not reproduce.

Watch

Extended reading notes

Core claim

The paper's central claim is that its CNN-based U-Net—built from Conv2dBlocks, max-pooling, Conv2DTranspose upsampling, and skip connections, with dropout at rate 0.07 and the Adam optimizer—can segment satellite landform images into accurate per-pixel class masks. Trained with binary cross-entropy on 5,000 preprocessed 512-by-512 images, the model achieves a Dice coefficient of 69.62% and a Model Accuracy of 90.53%; the paper describes the accuracy figure as the percentage of correctly classified instances during the training phase, and interprets the Dice value as a close pixel-level overlap with ground truth. It claims this performance is superior to the comparative models in its Table 1, which lists D-RESUNET at 62.93%, SiU-Net between 59.50% and 61.10%, and DRINet between 83.42% and 96.57%.

Load-bearing premise

The reported 69.62% Dice and 90.53% accuracy support the paper's claims only if they were measured on a held-out test set and if the comparison models were evaluated on the same landform data; the paper specifies neither condition.

Editorial extensions

If this is right

  • A U-Net trained on thousands of landform images can output segmentation masks that separate forests, deserts, mountains, and water at the pixel level.
  • A Dice coefficient of 69.62% would make the model usable for landform mapping tasks that tolerate moderate boundary overlap, such as regional land-use surveys.
  • The paper argues that scaling to much larger image collections would make predictions nearly indistinguishable from ground truth.
  • The same encoder-decoder recipe with dropout and Adam is adaptable to other segmentation tasks in satellite imagery, including cloud detection, water-body mapping, and building extraction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Extension: Because the paper describes Model Accuracy as a training-phase statistic, the 90.53% figure is not by itself evidence of generalization; a held-out test Dice would be the decisive number.
  • Extension: The comparison table mixes tasks and datasets (DRINet comes from medical-image segmentation), so the claim that the proposed model outperforms the listed methods would need a like-for-like benchmark on the same landform test set to be quantifiable.
  • Extension: The same U-Net configuration could be run on public satellite-imagery segmentation benchmarks to test whether the 69.62% Dice transfers to other landform datasets.
  • Extension: A small sweep around the chosen hyperparameters (dropout 0.07, Adam defaults) would show how sensitive the reported Dice is to these choices, which the paper leaves unexplored.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper applies a standard U-Net architecture with dropout, batch normalization, and the Adam optimizer to the semantic segmentation of landforms in 5,000 preprocessed 512×512 satellite images. It reports a Dice coefficient of 69.62% and a 'Model Accuracy' of 90.53%, and it claims that the proposed CNN-based U-Net outperforms comparative models. The manuscript includes descriptions of the U-Net architecture, the Adam optimizer, binary cross-entropy loss, and the Dice coefficient, followed by a results section and a comparison table.

Significance. If the reported performance were measured on a held-out test set and compared against models evaluated on the same data, the paper could serve as a useful empirical baseline for U-Net-based landform segmentation. However, the manuscript does not provide a valid evaluation protocol, does not release data or code, and its own comparison table contradicts the 'outperforms' claim. The paper makes no verifiable algorithmic contribution beyond applying a standard architecture with a dropout rate of 0.07. Its useful aspects are the applied problem setting and the use of standard segmentation metrics, but these are not enough to establish scientific significance without reproducible, out-of-sample evidence.

major comments (4)
  1. [Section 5, 'Experimentations and Results'] The paper defines 'Model Accuracy' as 'the percentage of correctly classified instances during the training phase' and states that the model 'generalize[s] well to seen examples.' No train/validation/test split is described anywhere in the manuscript. Therefore, the reported Dice coefficient of 69.62% and accuracy of 90.53% are in-sample quantities that do not support any claim about generalization to unseen satellite imagery. This invalidates the paper's central performance claim.
  2. [Section 5, Table 1 and accompanying text] Table 1 lists the proposed U-Net at 69.62%, D-RESUNET at 62.93%, DRINet at 83.42–96.57%, and SiU-Net at 59.50–61.10%. The text then asserts that the proposed model 'outperforms comparative models.' That statement is contradicted by the table itself, since the entire DRINet range lies above 69.62%. In addition, the comparators are taken from different datasets and tasks (DRINet on medical images, SiU-Net on building extraction), so the numbers are not commensurable. Both the internal contradiction and the dataset mismatch undermine the comparison.
  3. [Section 4, Equation (3)] Equation (3), presented as the Adam update rule, is algebraically incorrect. It writes θ_t = (θ_{t−1} − α) * m_t / (√v_t + ε), whereas the Adam algorithm subtracts α times the moment ratio from θ_{t−1}. The equation also omits the bias-correction terms present in the standard Adam formulation. While this is not the central issue, it is a concrete technical error in a stated mathematical formula.
  4. [General reproducibility] The manuscript does not provide the dataset source, a train/test split, the number of classes, the filter sizes, the learning rate, the batch size, the number of epochs, or any code or link to a public implementation. The only explicit hyperparameter given is a dropout rate of 0.07. Without these details, the experiments cannot be reproduced or independently verified, and the claimed performance cannot be checked.
minor comments (7)
  1. [Section 3 heading] The section title 'U-Net Architechture' contains a misspelling; it should be 'U-Net Architecture.'
  2. [Section 6, Conclusion] The conclusion says 'retaining spacial context'; the correct term is 'spatial context.'
  3. [Section 5, first paragraph] The dataset description is ambiguous: it first says 'a comprehensive dataset of 5,000 preprocessed satellite landform images,' then later says 'training the model on 5000 images of each of the segmented maps and landform maps.' It is unclear whether the dataset contains 5,000 images in total or 5,000 per map type, and whether these are separate images or triplets of input, height, and segmentation maps.
  4. [Section 5, Figure 3] Figure 3 is referenced as 'Performance Metrics of the U-Net Model' but the caption provides no explanation of the axes or whether the plotted values come from training or validation data; the figure appears to be a screenshot of a training history rather than a test-set metric.
  5. [Section 4, text around 'GiveMeUnet'] The implementation description mentions a function named 'GiveMeUnet' and unspecified 'specified filter sizes' and 'dropout rates,' but no concrete values are given, and the informal function name is unsuitable for a formal manuscript.
  6. [Section 4, paragraph on Dice coefficient] The statement that 'the Dice coefficient stands out as a superior metric compared to Intersection over Union (IoU) or accuracy' is presented as a general fact without supporting evidence specific to this landform segmentation task; this is an unsubstantiated assertion.
  7. [Section 4, Equation (1)] Equation (1), 'U − Net(x) = F(F(F(F(x))) −> ... −> F(x))', is not a well-defined mathematical expression; the recursion and the arrow notation are unclear and should be either formalized or removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model, loss, and metrics are standard external definitions, and the paper's weaknesses are evaluation-protocol soundness issues, not circular reasoning.

full rationale

The paper's derivation chain does not reduce to its inputs. The U-Net architecture, sigmoid activation, Adam optimizer, binary cross-entropy (Eq. 4), and Dice coefficient (Eq. 5) are standard external definitions and are not fitted to the reported outcome. Equation (1) is only a generic statement of recursive composition; it makes no substantive claim that is later used as evidence. The central quantitative claim, a Dice coefficient of 69.62 percent and a Model Accuracy of 90.53 percent, is an empirical report, not a derived prediction. Even if the evaluation protocol is unclear—the paper defines Model Accuracy as 'the percentage of correctly classified instances during the training phase' and never describes a held-out train/test split—that is a soundness or validity deficiency, not circularity. The comparison table is internally inconsistent: DRINet's reported lower bound of 83.42 percent exceeds the proposed model's 69.62 percent, while the text says the proposed model outperforms comparators; this is a factual and evidentiary weakness, not a reduction of a prediction to a fitted input. The only self-citation, reference [7], is to prior work on machine-learning techniques and routing in 5G and 6G networks; it is cited as background for the importance of satellite imagery and is not load-bearing for the segmentation result. No uniqueness theorem, ansatz-smuggling citation, or fitted-parameter-renamed-as-prediction pattern appears. The paper's conclusion also acknowledges limitations such as dependence on labeled data and sensitivity to hyperparameters, which are honest caveats rather than circular moves. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper relies on standard deep learning machinery and on an unverified private dataset. The central numerical claims depend on unreported hyperparameters and on the implicit assumption that metrics were computed on held-out data, which the text contradicts by calling accuracy a training-phase measure.

free parameters (4)
  • dropout rate = 0.07
    Chosen by hand in Section 4; the paper gives no search or ablation showing this value is optimal.
  • U-Net filter sizes = not reported
    Section 4 says GiveMeUnet uses "specified filter sizes" but never lists them.
  • Adam hyperparameters (learning rate, beta1, beta2, epsilon) = not reported
    Section 4 invokes Adam without values; the update equation is also written incorrectly.
  • train/validation split ratio = not reported
    Section 5 reports aggregated metrics without stating how the 5,000 images were divided.
assumptions (4)
  • domain assumption U-Net encoder-decoder with skip connections is an appropriate model for landform segmentation
    Invoked throughout Sections 3 and 4; the paper provides no comparison to other segmentation models on the same data.
  • domain assumption The 5,000 preprocessed 512x512 images and their segmentation maps are representative and correctly labeled
    Section 5 states the dataset without provenance, annotation protocol, or quality control.
  • ad hoc to paper Dice coefficient is a superior metric to IoU or accuracy for this task
    Section 4, paragraph after Equation (5), asserts this without evidence and uses it to justify reporting only Dice and accuracy.
  • ad hoc to paper Model Accuracy computed during training indicates generalization
    Section 5 calls a 90.53% training-phase accuracy evidence of generalization, which conflates fitting with prediction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Convolutional Neural Network Segmentation for Satellite Imagery Data to Identify Landforms Using U-Net Architecture." pith.science (2026). https://pith.science/paper/QMX6UXWU

@misc{pith2026250205476,
  author       = {Pith},
  title        = {Pith review of: Convolutional Neural Network Segmentation for Satellite Imagery Data to Identify Landforms Using U-Net Architecture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QMX6UXWU}},
  note         = {Machine review of arXiv:2502.05476}
}
read the original abstract

This study demonstrates a novel use of the U-Net architecture in the field of semantic segmentation to detect landforms using preprocessed satellite imagery. The study applies the U-Net model for effective feature extraction by using Convolutional Neural Network (CNN) segmentation techniques. Dropout is strategically used for regularization to improve the model's perseverance, and the Adam optimizer is used for effective training. The study thoroughly assesses the performance of the U-Net architecture utilizing a large sample of preprocessed satellite topographical images. The model excels in semantic segmentation tasks, displaying high-resolution outputs, quick feature extraction, and flexibility to a wide range of applications. The findings highlight the U-Net architecture's substantial contribution to the advancement of machine learning and image processing technologies. The U-Net approach, which emphasizes pixel-wise categorization and comprehensive segmentation map production, is helpful in practical applications such as autonomous driving, disaster management, and land use planning. This study not only investigates the complexities of U-Net architecture for semantic segmentation, but also highlights its real-world applications in image classification, analysis, and landform identification. The study demonstrates the U-Net model's key significance in influencing the environment of modern technology.

Figures

Figures reproduced from arXiv: 2502.05476 by the authors.

Figure 1
Figure 1. U-Net Model Flow Diagram Let U-Net(x) represent the U-Net architecture's forward pass on input x, and F be the set of operations applied in each block. The equation can be roughly summarized as: 𝑈 −𝑁𝑒𝑡(𝑥) = 𝐹(𝐹(𝐹(𝐹(𝑥)))−> . . .−> 𝐹(𝑥)) (1) Equation (1) highlights the recursive and hierarchical nature of the U-Net architecture, where the input undergoes a series of operations (F) at multiple levels, with skip connect… view at source ↗
Figure 2
Figure 2. Comparing Raw Landform Data, the Actual Landform Mask, and the Predicted Landform Mask The acquired findings provide convincing proof of the U-Net model's competence in image segmentation. It effectively parses the subtle features in the raw landform images and produces a segmented result with an impressive level of accuracy. This potential is seen when the predicted images are compared to the real landform mask, re… view at source ↗
Figure 3
Figure 3. Performance Metrics of the U-Net Model The model achieves a commendable Dice Coefficient of 69.62% and a Model Accuracy of 90.53% (see [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages

  1. [1]

    Electron Markets 31(6), 685–695 (2021)

    Janiesch, C., Zschech, P., Heinrich, K.: Machine learning and deep learning. Electron Markets 31(6), 685–695 (2021). https://doi.org/10.1007/s12525-021-00475-2

    doi:10.1007/s12525-021-00475-2 2021

  2. [2]

    Pattern Recognition Letters 141, 61–67 (2021)

    Wang, P., Fan, E., Wang, P.: Comparative analysis of image classification algorithms based on traditional machine learning and deep learning. Pattern Recognition Letters 141, 61–67 (2021). https://doi.org/10.1016/j.patrec.2020.07.042

    doi:10.1016/j.patrec.2020.07.042 2021

  3. [3]

    Procedia Technology 24, 1366 –1373 (2016)

    Vijayan, A., Kareem, S., Kizhakkethottam, J.J.: Face Recognition Across Gender Trans- formation Using SVM Classifier. Procedia Technology 24, 1366 –1373 (2016). https://doi.org/10.1016/j.protcy.2016.05.150

    work page doi:10.1016/j.protcy.2016.05.150 2016

  4. [4]

    2018 Baltic URSI Symposium (URSI), Poznan, Poland, pp

    Kozlowski, P., Walas, K.: Deep neural networks for landform recognition task. 2018 Baltic URSI Symposium (URSI), Poznan, Poland, pp. 283 –286 (2018). https://doi.org/10.23919/URSI.2018.8406736

    work page arXiv 2018

  5. [5]

    2016 IEEE 19th Interna- tional Conference on Intelligent Transportation Systems (ITSC), Rio de Janeiro, Brazil, pp

    Kutila, M., Pyykönen, P., Ritter, W., Sawade, O., Schäufele, B.: Automotive LIDAR sensor development scenarios for harsh weather conditions. 2016 IEEE 19th Interna- tional Conference on Intelligent Transportation Systems (ITSC), Rio de Janeiro, Brazil, pp. 265–270 (2016). https://doi.org/10.1109/ITSC.2016.7795565

    work page arXiv 2016

  6. [6]

    IEEE Transactions on Sys- tems, Man, and Cybernetics 8(4), 237 –247 (1978)

    Sklansky, J.: Image Segmentation and Feature Extraction. IEEE Transactions on Sys- tems, Man, and Cybernetics 8(4), 237 –247 (1978). https://doi.org/10.1109/TSMC.1978.4309944

    work page arXiv 1978

  7. [7]

    2023 7th International Conference on Trends in Electronics and Informatics (ICOEI), Tirun elveli, India, pp

    Goswami, M., Panda, N., Mohanty, S., Pattnaik, P.K.: Machine Learning Techniques and Routing Protocols in 5G and 6G Mobile Network Communication System - An Overview. 2023 7th International Conference on Trends in Electronics and Informatics (ICOEI), Tirun elveli, India, pp. 1094 –1101 (2023). https://doi.org/10.1109/ICOEI56765.2023.10125697

    work page arXiv 2023

  8. [8]

    In: 2021 28th Conference of Open Innovations As- sociation (FRUCT), pp

    Bagaev, S., & Medvedeva, E.: Segmentation of satellite images of the earth’s surface using neural network technologies. In: 2021 28th Conference of Open Innovations As- sociation (FRUCT), pp. 15–21 (2021)

    work page 2021

Show all 24 references
  1. [9]

    Scientific Reports, vol

    Kattenborn, T., Eichel, J., & Fassnacht, F.: Convolutional neural networks enable effi- cient, accurate and fine -grained segmentation of plant species and communities from high-resolution UAV imagery. Scientific Reports, vol. 9 (2019)

    2019

  2. [10]

    In: 2020 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineer- ing (EIConRus), pp

    Tiurin, A., Vorobiev, M., Lisov, O., Andrianov, A., & Yanakova, E.: An effective algo- rithm for analysis and processing of satellite images for semantic segmentation. In: 2020 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineer- ing (EIConRus), ...

    2020

  3. [11]

    In: 2019 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), pp

    Gonzales, C., & Sakla, W.: Semantic segmentation of clouds in satellite imagery using deep pre-trained U-nets. In: 2019 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), pp. 1–7 (2019). 14

    2019

  4. [12]

    In: IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, pp

    Liu, Z., Feng, R., Wang, L., Zhong, Y., & Cao, L.: D-resunet: Resunet and dilated con- volution for high resolution satellite imagery road extraction. In: IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, pp. 3927 –3930 (2019)

    2019

  5. [13]

    IEEE Transactions on Medical Imaging, vol

    Chen, L., Bentley, P., Mori, K., Misawa, K., Fujiwara, M., & Rueckert, D.: Drinet for medical image segmentation. IEEE Transactions on Medical Imaging, vol. 37, pp. 2453– 2462 (2018)

    2018

  6. [14]

    In: IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, pp

    Grosgeorge, D., Arbelot, M., Goupilleau, A., Ceillier, T., & Allioux, R.: Concurrent segmentation and object detection CNNs for aircraft detection and identification in sat- ellite images. In: IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, pp. 2...

    2020

  7. [15]

    In: 2019 Systems of Signal Synchronization, Generating and Processing in Telecommuni- cations (SYNCHROINFO), pp

    Sedov, A., Khryashchev, V., Larionov, R., & Ostrovskaya, A.: Loss function selection in a problem of satellite image segmentation using convolutional neural network. In: 2019 Systems of Signal Synchronization, Generating and Processing in Telecommuni- cations (SYNCHROINFO), pp...

    2019

  8. [16]

    ArXiv, vol

    Gonzalez, J., Bhowmick, D., Beltrán, C., Sankaran, K., & Bengio, Y.: Applying knowledge transfer for water body segmentation in Peru. ArXiv, vol. abs/1912.00957 (2019)

    2019 arXiv

  9. [17]

    IEEE Transactions on Geoscience and Remote Sensing, vol

    Ji, S., Wei, S., & Lu, M.: Fully convolutional networks for multisource building extrac- tion from an open aerial and satellite imagery dataset. IEEE Transactions on Geoscience and Remote Sensing, vol. 57, pp. 574–586 (2019)

    2019

  10. [18]

    Neural Networks 121, 74 –87 (2020)

    Ibtehaz, N., Rahman, M.S.: MultiResUNet: Rethinking the U-Net architecture for mul- timodal biomedical image segmentation. Neural Networks 121, 74 –87 (2020). https://doi.org/10.1016/j.neunet.2019.08.025

    2020 doi

  11. [19]

    Journal of Healthcare Engineering, vol

    Yin, X.-X., Sun, L., Fu, Y., Lu, R., Zhang, Y.: U-Net-Based Medical Image Segmenta- tion. Journal of Healthcare Engineering, vol. 2022, Article ID 4189781, 16 pages (2022). https://doi.org/10.1155/2022/4189781

    2022 doi

  12. [20]

    IEEE Access 7, 44247 –44257 (2019)

    Weng, Y., Zhou, T., Li, Y., Qiu, X.: NAS-Unet: Neural Architecture Search for Medical Image Segmentation. IEEE Access 7, 44247 –44257 (2019). https://doi.org/10.1109/ACCESS.2019.2908991

    2019

  13. [21]

    Pattern Recognition 44(4), 777 –787 (2011)

    Wang, X.-Y., Wang, T., Bu, J.: Color image segmentation using pixel-wise support vec- tor machine classification. Pattern Recognition 44(4), 777 –787 (2011). https://doi.org/10.1016/j.patcog.2010.08.008

    2011 doi

  14. [22]

    In: Dash, S., Acharya, B., Mittal, M., Abraham, A., Kelemen, A

    Kumar, A., Sarkar, S., Pradhan, C.: Malaria Disease Detection Using CNN Technique with SGD, RMSprop and ADAM Optimizers. In: Dash, S., Acharya, B., Mittal, M., Abraham, A., Kelemen, A. (eds.) Deep Learning Techniques for Biomedical and Health Informatics, Studies in Big Data, ...

    2020 doi

  15. [23]

    2019 41st Annual International Conference of the IEEE Engineering in Medi- cine and Biology Society (EMBC), Berlin, Germany, pp

    Hasan, S.M.K., Linte, C.A.: U-NetPlus: A Modified Encoder-Decoder U-Net Architec- ture for Semantic and Instance Segmentation of Surgical Instruments from Laparoscopic Images. 2019 41st Annual International Conference of the IEEE Engineering in Medi- cine and Biology Society (...

    2019

  16. [24]

    Applied Soft Computing 126, 109297 (2022)

    Cheng, J., Tian, S., Yu, L., Liu, S., Wang, C., Ren, Y., Lu, H., Zhu, M.: DDU-Net: A dual dense U-structure network for medical image segmentation. Applied Soft Computing 126, 109297 (2022). https://doi.org/10.1016/j.asoc.2022.109297

    2022

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.