Pith. sign in

REVIEW 4 major objections 6 minor 11 references

Residual Connection Networks in Medical Image Processing: Exploration of ResUnet++ Model Driven by Human Computer Interaction

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read The paper introduces ResUnet++, a residual-block U-Net hybrid for brain tumor MRI segmentation, and reports a 98.17% Jaccard/IoU on the LGG dataset while framing the tool as clinician-interactive.

desk verdict The paper's central 98.17% IoU claim is contradicted by its own validation loss, the architecture is a re-branding of existing work, and the HCI component is unevaluated; reject outright. read the letter →

arxiv 2412.20709 v1 pith:LA3OG3WO submitted 2024-12-30 eess.IV cs.CV

classification eess.IVcs.CV MSC 68T0768U1092C55
keywords ResUnet++braintumorsegmentationmedicalimageresiduallearningU-NetJaccardlosshuman-computerinteractionLGGdataset
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 proposes ResUnet++, a hybrid model that inserts ResNet-style residual blocks into a U-Net encoder-decoder and adds ASPP multi-scale pooling plus attention. The authors claim that on the LGG brain-tumor MRI dataset this model reaches a Jaccard/IoU of $98.17\%$, which they call both a Jaccard loss and an IoU index. They further argue that human-computer interaction features, real-time visualization and feedback, make the segmentation tool practical for clinicians. If the performance number holds, ResUnet++ would be a near-exact tumor segmenter with a built-in path to clinical interaction.

What carries the argument

The residual block is the mechanism that carries the architecture: a shortcut connection adds the block's input to its transformed output, so features survive the down- and upsampling depths. The Jaccard loss is the identity that produces the headline number, computing intersection over union between predicted and ground-truth tumor regions. Around these sit the ASPP module for multi-scale context and an attention mechanism for suppressing irrelevant regions.

What would settle it

Re-run ResUnet++ on the LGG dataset with an explicit held-out test split and at least three random seeds, computing the mean test IoU; if it falls well below $0.9817$, or if $1 - \text{Jaccard loss}$ does not match the reported validation loss of $0.1718$, the paper's core performance claim does not hold.

Watch

Extended reading notes

Core claim

ResUnet++'s downsampling path alternates convolution, batch normalization, ReLU, max pooling, and residual blocks, growing from 16 to 256 feature channels; the upsampling path mirrors this with up-convolutions, concatenation, and residual blocks, ending with a $1\times1\times1$ convolution to a single output channel. The model is trained for 60 epochs with the NAdam optimizer, learning-rate decay, early stopping, and Jaccard loss. The paper's central reported result is a single number: $\mathrm{IoU} = 98.17\%$ on the LGG Segmentation Dataset, which it interprets as good model fit and uses to motivate real-time clinician interaction. The authors also report training and validation losses of 0.0764 and 0.1718 after 60 epochs.

Load-bearing premise

The headline number only means what it seems to mean if $98.17\%$ is a Jaccard/IoU score computed on an independent test split, which the paper never defines; the same number is also called a loss, and the reported training/validation losses do not line up with it.

Editorial extensions

If this is right

  • If the 98.17% Jaccard/IoU holds on held-out data, ResUnet++ could generate whole-tumor contours accurate enough to reduce manual delineation effort in lower-grade glioma cases.
  • The HCI layer would give clinicians real-time visualization and correction of predicted tumor regions, potentially increasing trust in automated segmentations.
  • The residual-plus-ASPP-plus-attention recipe would transfer to other MRI segmentation targets with modest changes to the encoder depth.
  • Using Jaccard loss alone would remove the need for hand-tuned class weights on imbalanced tumor masks.

Reading between the lines

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

  • A consequence the authors do not draw: their HCI claim is about usability, so it cannot be tested from the reported numbers; a small usability study with radiologists would be the direct next experiment.
  • A testable extension is to report Dice, per-fold variance, and a defined test split on LGG; that would show whether 98.17% is stable or a single training run.
  • The text uses Jaccard loss, Jaccard index, and IoU for the same number; auditing which quantity was actually computed is the first step before comparing with other models.
  • A direct comparison against U-Net, UNet++, and nnU-Net under the same split would place ResUnet++ in context, but the paper provides no baseline comparison.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes ResUnet++, a hybrid U-Net/ResNet architecture with residual blocks in both downsampling and upsampling paths, ASPP, and an attention mechanism, for brain tumor segmentation on the LGG MRI dataset. It claims that the model achieves a 98.17% Jaccard/IoU score and that the approach integrates human-computer interaction to improve clinical workflow. The manuscript includes a literature survey, architecture description, experimental section, and conclusion. The central empirical claim is the reported 98.17% performance, and the HCI integration is presented as the distinctive contribution alongside the architecture.

Significance. If the performance and HCI claims were substantiated, the paper would offer a useful benchmark-style contribution to medical image segmentation and a pathway for interactive clinical tools. However, the manuscript does not supply the evidence needed to support these claims: the headline metric is internally inconsistent with the reported loss values, no independent test set is defined, and the HCI component is never implemented or evaluated. The main positive features are the understandable motivation for residual connections in segmentation and the reproducible problem formulation, but the missing validation and contradictory numbers undermine the contribution. No code, reproducible experiments, or formal evaluations are provided, so the significance of the work as a scientific contribution is currently minimal.

major comments (4)
  1. [Abstract, §3.2, §4.2, §5] The central performance claim is internally inconsistent with the reported loss values. If the model is trained with Jaccard Loss, as stated in Section 3.2, then a validation loss of 0.1718 reported in Section 5 corresponds to an IoU of roughly 82.8%, not 98.17%. Conversely, an IoU of 98.17% would require a validation loss near 0.0183. Moreover, the phrase 'Jaccard Loss of 98.17%' in the Abstract is a category error: a loss of 98.17% would indicate near-zero overlap, not high segmentation quality. The authors must reconcile the metric definition, the exact number, and the subset on which it is computed.
  2. [§3.2, §4.2] The manuscript never defines a train/validation/test split. Section 4.1 mentions the LGG dataset is used for training, validation, and testing, but no split ratio, patient-level separation, or evaluation protocol is given. The reported 98.17% IoU therefore cannot be attributed to independent test data, and hyperparameters (learning rate, epsilon, number of epochs) appear to be selected using the same data. The authors must provide a clear data split and report metrics on a held-out test set, or revise the claim accordingly.
  3. [Abstract, §1, §3-§5] The claimed HCI contribution is not evaluated anywhere in the paper. The abstract and introduction state that ResUnet++ provides real-time feedback and clinician interaction, but Sections 3 through 5 contain no user study, interface description, usability metrics, or clinician-in-the-loop experiments. The phrase 'HCI principles are integrated' in Section 1 is unsupported by any methodological or experimental detail, so the HCI narrative is only motivational.
  4. [§3.2 (two headings)] The model description is inconsistent. The first Section 3.2 says the network includes an ASPP module and attention mechanism, but Figures 1 and 2, which are described as the downsampling and upsampling sections, show no such components. The same section states the model uses the Adam optimizer with learning rate 0.05 and epsilon 0.1, while the second Section 3.2 (NAdam optimizer and learning rate) says NAdam is chosen; the paper never clarifies which optimizer was actually used. Each architectural component and hyperparameter should be specified unambiguously.
minor comments (6)
  1. [Fig. 4] Figure 4's caption 'The change of accuary for resunet++' contains a typo ('accuary') and does not identify what accuracy or loss is plotted or on which data split; the axes are also unlabeled.
  2. [Fig. 5 caption] The caption for Fig. 5 references 'figure10' rather than 'Fig. 5' and should be corrected.
  3. [§3.2] The text describing the upsampling path says 'For downsampling, a 2×2 upsampling layer doubles the image size'; this appears to be a typo for 'upsampling' and should be fixed.
  4. [§4.2] The experimental results section reports only a single IoU value (98.17%) with no error bars, no comparison to baseline or state-of-the-art methods, and no statistical significance testing; the reader cannot assess whether this value is representative.
  5. [§2] The related-work section does not cite any prior ResUnet++ literature; since ResUnet++ is an existing architecture in medical image segmentation, the novelty of the proposed model should be clarified with explicit comparison.
  6. [References] The reference list would benefit from standardizing entries (e.g., complete volume/page information for the IEEE journals) and from checking that all in-text citations match the bibliography; some entries lack full bibliographic details.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported 98.17% IoU is the training objective on an unspecified split, presented as an independent prediction.

  1. fitted input called prediction [Abstract, Section 3.2, Section 4.2, Section 5]
    "We replace traditional loss functions with Jaccard Loss, which calculates the ratio of intersection to union between predicted and true regions. ... Our IoU index reaches 98.17%, indicating a good model fit. ... After 60 training epochs, the training set loss rate is 0.0764, and the validation set loss rate is 0.1718, demonstrating excellent performance."

    The model is trained by minimizing Jaccard Loss, which is exactly 1 minus IoU, so high IoU on the fitting data is forced by the optimization objective. The paper never defines a train/validation/test split, and Section 4.2 interprets the 98.17% IoU as 'a good model fit' rather than as an independent test result. Thus the central reported 'prediction' reduces to the training objective evaluated on the data used for fitting and model selection. Additionally, the paper's own validation loss of 0.1718 corresponds to an IoU of roughly 82.8%, not 98.17%, so the headline number is not even consistent with the reported validation result.

full rationale

The paper contains no self-citations and no imported uniqueness or ansatz argument, so the circularity here is not a citation-chain problem. Instead, the central empirical claim — 98.17% Jaccard/IoU performance on the LGG dataset — is presented in the Abstract and Section 4.2 as the paper's main result, but the paper never specifies on which subset of data this number was computed. The loss used for training is Jaccard Loss, and IoU is the same intersection-over-union quantity, so reporting high IoU on a training or model-selection split is guaranteed by optimization rather than being a predictive finding. The conclusion's validation loss of 0.1718 directly contradicts the claimed 98.17% IoU under the standard Jaccard-loss definition, confirming that the headline number is not an independent validation result. There is also a category error in the phrase 'Jaccard Loss of 98.17%', since a loss of 98.17% would indicate poor overlap. Because the central contribution reduces to a fitted value presented as a prediction, the circularity score is 6.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central performance claim rests entirely on an unspecified evaluation protocol and on standard deep-learning assumptions; no new entity is introduced. The reported metric is likely a training or validation fit, making the result self-referential.

free parameters (7)
  • Learning rate = 0.05
    Adam optimizer specified in Section 3.2; extremely high for segmentation, and no schedule reported despite Section 3.3 claiming decay.
  • Optimizer choice = Adam or NAdam, contradictory
    Section 3.2 says Adam; Section 3.3 says NAdam was chosen. Both cannot be the actual configuration.
  • Epsilon = 0.1
    Listed in Section 3.2 without justification.
  • Training epochs = 60
    Fixed in Sections 3.2 and 5 without an epoch curve or early stopping analysis.
  • Filter widths = 16, 32, 64, 128, 256
    Architecture widths chosen in Section 3.2 with no ablation study.
  • Input size = 256x256x3
    Set in Section 3.2, with no comparison to other input resolutions.
  • Jaccard loss function = Jaccard (IoU-based)
    Chosen in place of cross-entropy in Section 3.2, but no comparison to other loss functions is shown.
assumptions (4)
  • domain assumption Ground-truth masks in the LGG dataset are accurate and unambiguous.
    The model is trained and evaluated against these masks without any label-quality audit (Section 4.1).
  • domain assumption The reported metric is computed on a held-out test set.
    The paper never states a train/validation/test split, and the conclusion reports training and validation losses that do not match the claimed Jaccard index (Sections 4.2 and 5).
  • domain assumption Residual connections, ASPP, and attention improve segmentation as claimed.
    Architectural choices are asserted from literature without ablations or comparisons on this dataset (Section 3.2).
  • standard math Jaccard loss equals 1 minus IoU.
    Used implicitly when presenting 98.17% as both 'Jaccard Loss' and 'IoU index' (Abstract and Section 4.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Residual Connection Networks in Medical Image Processing: Exploration of ResUnet++ Model Driven by Human Computer Interaction." pith.science (2026). https://pith.science/paper/LA3OG3WO

@misc{pith2026241220709,
  author       = {Pith},
  title        = {Pith review of: Residual Connection Networks in Medical Image Processing: Exploration of ResUnet++ Model Driven by Human Computer Interaction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LA3OG3WO}},
  note         = {Machine review of arXiv:2412.20709}
}
read the original abstract

Accurate identification and localisation of brain tumours from medical images remain challenging due to tumour variability and structural complexity. Convolutional Neural Networks (CNNs), particularly ResNet and Unet, have made significant progress in medical image processing, offering robust capabilities for image segmentation. However, limited research has explored their integration with human-computer interaction (HCI) to enhance usability, interpretability, and clinical applicability. This paper introduces ResUnet++, an advanced hybrid model combining ResNet and Unet++, designed to improve tumour detection and localisation while fostering seamless interaction between clinicians and medical imaging systems. ResUnet++ integrates residual blocks in both the downsampling and upsampling phases, ensuring critical image features are preserved. By incorporating HCI principles, the model provides intuitive, real-time feedback, enabling clinicians to visualise and interact with tumour localisation results effectively. This fosters informed decision-making and supports workflow efficiency in clinical settings. We evaluated ResUnet++ on the LGG Segmentation Dataset, achieving a Jaccard Loss of 98.17%. The results demonstrate its strong segmentation performance and potential for real-world applications. By bridging advanced medical imaging techniques with HCI, ResUnet++ offers a foundation for developing interactive diagnostic tools, improving clinician trust, decision accuracy, and patient outcomes, and advancing the integration of AI in healthcare workflows.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 9 canonical work pages

  1. [1]

    Deep Residual Learning for Image Recognition

    He, Kaiming , et al. "Deep Residual Learning for Image Recognition." IEEE (2016)

  2. [2]

    U -Net: Convolutional Networks for Biomedical Image Segmentation

    Ronneberger, Olaf , P. Fischer , and T. Brox . "U -Net: Convolutional Networks for Biomedical Image Segmentation." International Conference on Medical Image Computing and Computer - Assisted Intervention Springer International Publishing, 2015

  3. [3]

    RA-UNet: A hybrid deep attention-aware network to extract liver and tumor in CT scans

    Jin, Qiangguo , et al. "RA-UNet: A hybrid deep attention-aware network to extract liver and tumor in CT scans." (2018)

  4. [4]

    Attention U-Net: Learning Where to Look for the Pancreas

    Oktay, Ozan , et al. "Attention U-Net: Learning Where to Look for the Pancreas." (2018)

  5. [5]

    UNet 3+: A Full-Scale Connected UNet for Medical Image Segmentation

    Huang, Huimin , et al. "UNet 3+: A Full-Scale Connected UNet for Medical Image Segmentation." arXiv (2020)

  6. [6]

    Pyramid Scene Parsing Network

    Zhao, Hengshuang et al. “Pyramid Scene Parsing Network.” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016): 6230-6239

  7. [7]

    DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs

    Chen, Liang-Chieh et al. “DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs.” IEEE Transactions on Pattern Analysis and Machine Intelligence 40 (2016): 834-848

  8. [8]

    Rethinking Atrous Convolution for Semantic Image Segmentation

    Chen, Liang -Chieh et al. “Rethinking Atrous Convolution for Semantic Image Segmentation.” ArXiv abs/1706.05587 (2017): n. pag

Show all 11 references
  1. [9]

    Encoder -Decoder with Atrous Separable Convolution for Semantic Image Segmentation

    Chen, Liang -Chieh et al. “Encoder -Decoder with Atrous Separable Convolution for Semantic Image Segmentation.” European Conference on Computer Vision (2018)

  2. [10]

    MDD-UNet: Domain Adaptation for Medical Image Segmentation with Theoretical Guarantees, a Proof of Concept

    Munk, Asbjorn et al. “MDD-UNet: Domain Adaptation for Medical Image Segmentation with Theoretical Guarantees, a Proof of Concept.” NLDL (2023)

  3. [11]

    VM -UNet: Vision Mamba UNet for Medical Image Segmentation

    Ruan, Jiacheng and Suncheng Xiang. “VM -UNet: Vision Mamba UNet for Medical Image Segmentation.” arXiv abs/2402.02491 (2024): n. pag

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.