Pith. sign in

REVIEW 5 major objections 6 minor 22 references

Semi-Supervised Multi-Task Learning With Chest X-Ray Images

T0 review · 5 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a new KLTV loss lets a single adversarial U-Net jointly classify and segment chest X-rays from mostly unlabeled data, consistently outperforming TV and XETV baselines.

desk verdict The proposed KLTV loss is undefined for background pixels as written, so the paper's central empirical claim is unverifiable, but the multi-task architecture and problem setup are worth corrective review. read the letter →

arxiv 1908.03693 v2 pith:7KXIB2IZ submitted 2019-08-10 eess.IV cs.CV

Abdullah-Al-Zubaer Imran , Demetri Terzopoulos This is my paper · ORCID
classification eess.IVcs.CV
keywords semi-supervisedlearningmulti-taskchestX-rayimagesegmentationgenerativeadversarialnetworksKLdivergenceTverskylossattentionU-Net
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 single generative-adversarial model can learn to classify chest X-rays and segment lungs at the same time from mostly unlabeled data, and that joint training benefits from a specially designed segmentation loss. The loss, KLTV, combines absolute KL divergence with Tversky loss, and the model, APPAU-Net, pairs a pyramid progressive attention U-Net segmentor with an adversarial discriminator that doubles as a disease classifier. The paper reports that in semi-supervised multi-task experiments with only 10 percent labeled training data, the KLTV version consistently outperformed the TV and XETV versions across four chest X-ray datasets on both overlap and distance measures. The authors present this as a hypothesis that KLTV is more effective for generalizing multi-tasking models while remaining competitive in segmentation-only tasks. If correct, the practical payoff is a way to extract joint classification and segmentation from small, partially labeled medical image collections.

What carries the argument

The central object is APPAU-Net, composed of a segmentor and a discriminator. The segmentor is a pyramid progressive attention U-Net (PPAU-Net): a U-Net-like encoder-decoder that feeds the image at several scales into the encoder, generates side-output segmentations at different decoder resolutions, and uses attention gates to prune low-level features before they are merged with high-level ones. The discriminator is an n-plus-one-class CNN that receives either an image with its ground-truth mask or an image with the segmentor's predicted mask, assigning the first n classes to real disease labels and an extra class to synthetic pairs; this makes it both the adversarial critic and the classifier. The mechanism that carries the argument is the KLTV loss, defined as a weighted sum of an absolute KL-divergence term between ground-truth and predicted probability maps and a Tversky loss that separately weights false positives and false negatives, plus a small adversarial term. The paper's case is that this combination yields faster convergence and better segmentation than Tversky alone or cross-entropy plus Tversky, especially when the multi-task objective competes with the segmentation objective.

What would settle it

Run the same four-dataset comparison with KLTV versus TV while setting the KL weight to zero, or inspect the training loop to see whether unlabeled images ever pass through the segmentor; if KLTV no longer wins on overlap and distance, or if every training minibatch turns out to be fully labeled, the central semi-supervised advantage claimed for the method is not supported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is the stated hypothesis: KLTV can be more effective for generalizing multi-tasking models while being competitive in segmentation-only tasks. Concretely, the APPAU-Net model trained with the proposed KLTV loss consistently outperformed the same model trained with TV or XETV losses in both overlap and distance measures, across the MCX, SCX, JCX, and CCX chest X-ray datasets, in a semi-supervised multi-task setting with 10 percent labeled data. In segmentation-only comparisons, the KLTV-trained PPAU-Net matched or exceeded earlier U-Net variants on Dice similarity and Hausdorff distance for most datasets. The paper also claims that this is the first multi-task learning approach to chest X-ray analysis, and that the adversarial discriminator performs disease classification and authenticity checking simultaneously.

Load-bearing premise

The training procedure as written assumes every minibatch has ground-truth labels and masks and feeds them to the discriminator as real pairs, even though the experiments say only 10 percent of the training data is labeled; the mechanism for using the remaining unlabeled images is not stated in the paper.

Editorial extensions

If this is right

  • A semi-supervised multi-task model can reach segmentation performance close to fully supervised single-task models while also outputting disease classifications, reducing the annotation burden for chest X-ray analysis.
  • KLTV is a viable drop-in segmentation loss for small-data medical imaging, since the segmentation-only results remain competitive with earlier U-Net variants.
  • The discriminator's dual role as classifier and adversarial critic means a single extra network can serve both purposes without a separate classification head.
  • Combining multiple chest X-ray sources into one training set supports 3-class classification, suggesting the model can absorb heterogeneous data.

Reading between the lines

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

  • The consistent gains of KLTV over TV hint that the absolute KL term acts as a distributional regularizer, preventing the segmentor's feature maps from being pulled away by the classification task; testing KLTV on other small medical image multi-task setups would reveal whether the benefit is general.
  • Because Algorithm 1 in the appendix appears to feed every minibatch as an image-label pair, the mechanism by which the 90 percent unlabeled images enter training is unclear; an implementation that actually runs unlabeled images through the segmentor would determine whether the reported gains come from semi-supervision or from the KLTV loss alone.
  • The CCX 3-class result suggests multi-task training may help rare-class recognition by sharing features with segmentation; comparing APPAU-Net-KLTV to a classifier-only model with the same backbone would isolate that effect.
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

5 major / 6 minor

Summary. The paper proposes APPAU-Net, a generative multi-task model for joint classification and segmentation of chest X-ray images, trained in a semi-supervised manner with 10% labeled data. A new loss function, KLTV, is introduced that combines absolute KL divergence with the Tversky loss. The authors report experiments on four chest X-ray datasets (MCX, SCX, JCX, and a combined CCX) and claim that the APPAU-Net with KLTV consistently outperforms the same model with TV and XETV losses in both overlap and distance metrics for segmentation while remaining competitive in classification. The paper also introduces a pyramid progressive attention U-Net architecture with deep supervision and adversarial training.

Significance. If the technical claims are correct, the work addresses a relevant and practical problem—semi-supervised multi-task learning for medical image analysis with scarce labeled data. The proposed architecture, APPAU-Net, with pyramid inputs and progressive attention-gated side outputs, is a sensible extension of existing U-Net variants. The KLTV loss, if properly defined, could be a useful contribution for handling class imbalance in segmentation. The authors provide a reasonably broad experimental comparison across multiple datasets and metrics. However, the manuscript as written contains a serious flaw in the definition of the proposed loss, an inconsistency in the description of the semi-supervised training procedure, and a lack of statistical validation; these issues must be resolved before the central claims can be accepted.

major comments (5)
  1. [Section 2.2, Eq. (4)] This is the most serious technical issue because it affects every reported KLTV result, including the central claim of the paper.
  2. [Appendix B, Algorithm 1 and Section 3] This issue bears directly on the claimed novelty of semi-supervised multi-task learning and must be resolved before the empirical results can be interpreted.
  3. [Section 3, Tables 1 and 2] This is required for the central comparative claim to be credible.
  4. [Appendix A, item 2] This concern affects the external validity of the experiments, though it is secondary to the loss and algorithm issues.
  5. [Abstract and Section 3] This is a claim that is not currently supported by the presented experiments.
minor comments (6)
  1. [Section 2.2, Eq. (4)] The notation for the subscripts in LSKL and LSTV is inconsistent: y_pl and y_pbar are used without an explicit definition of 'pl' and 'pbar' near their first use. Clarifying that 'pl' denotes the ROI/foreground label and 'pbar' the background label would improve readability.
  2. [Section 3, 'Implementation Details'] The learning rates are written as '1.0−5' and '1.0−4'; these should be typeset as 1e-5 and 1e-4 or 10^-5 and 10^-4 to avoid ambiguity.
  3. [Table 2] In the JCX rows, the Precision (PR) for classification is 0.000 for all three models. This suggests a degenerate classifier (likely predicting only the majority class). The authors should either explain this behavior or discuss why the classification task is not meaningful for this dataset; as is, the classification results for JCX are not informative.
  4. [Figure 3 and Figure 4] The visual comparisons show only a small number of example images. Adding more examples or providing quantitative zoomed views would make the claimed qualitative superiority of KLTV more convincing.
  5. [Section 2.2] There is a typo in the sentence 'weighs false negatives higher than than false positives'; remove the duplicated 'than'.
  6. [Section 1.1] The related work section is concise; referencing more recent multi-task learning and semi-supervised medical image analysis work would help place the contribution in context.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the KLTV advantage is an empirical comparison, not an identity or fitted-input prediction.

full rationale

The paper's central claim is empirical: the proposed KLTV loss is explicitly defined in Eq. (4) as a combination of absolute KL divergence and Tversky loss, and its reported advantage is established by comparing APPAU-Net-KLTV against APPAU-Net-TV and APPAU-Net-XETV in Tables 1 and 2. No equation or derivation in the paper reduces KLTV's measured performance to a fitted parameter, to a quantity defined in terms of the target result, or to a result imported solely from the authors' prior work. The self-citations (Refs. [8] and [15]) supply architectural ingredients—multi-adversarial variational autoencoder networks and progressive dense V-networks—but the KLTV comparison is evaluated on held-out test sets and against external baseline models, so those citations are not load-bearing for the main claim. The undefined KL term for background pixels and the tension between Algorithm 1's fully-labeled minibatch sampling and the 10%-labeled semi-supervised setup are correctness and consistency concerns, not circularity. No uniqueness theorem, ansatz, or renaming is used to force the conclusion, so the derivation chain is self-contained with respect to the claimed input-output relationship.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The central claim depends on a set of unreported loss weights and on the reliability of public ground-truth masks. The most consequential assumption is that the semi-supervised procedure actually works as described, because Algorithm 1 appears to use labels for every image. No new physical or theoretical entities are introduced.

free parameters (7)
  • a (KL weight in KLTV) = not reported
    Weight balancing absolute KL divergence against Tversky loss in Eq. (3); no value is given and it affects every segmentation result.
  • b (Tversky weight in KLTV) = not reported
    Weight balancing Tversky loss in Eq. (3); not reported.
  • c (adversarial loss weight) = not reported
    Weight on the segmentor adversarial loss in Eq. (3); described as small but no value is given.
  • alpha (false-positive penalty) = not reported
    Tversky denominator coefficient in Eq. (4); not specified.
  • beta (false-negative penalty) = not reported
    Tversky denominator coefficient in Eq. (4); not specified.
  • epsilon (numerical stabilizer) = not reported
    Small constant preventing division by zero in Eq. (4); value not specified.
  • side-output weights w_i = not reported
    Eq. (1) weights deep-supervision side outputs by resolution; no values are given.
assumptions (3)
  • domain assumption Public chest X-ray ground-truth masks from MCX, SCX, and JSRT are anatomically reliable and consistent across datasets.
    All segmentation supervision and evaluation rest on these masks; SCX images that did not agree with masks were removed in Appendix A, so the retained masks are assumed correct.
  • domain assumption For unlabeled training images, the segmentor's predicted masks are informative enough to train the discriminator and classifier.
    This is the load-bearing premise of the semi-supervised learning scheme in Section B and Algorithm 1; if predicted masks are too noisy, the unsupervised adversarial signal would not improve classification.
  • ad hoc to paper The absolute KL divergence term in L_SKL is well-defined for all pixel predictions.
    Eq. (4) contains log(y_pl / y_hat_pl) but no epsilon or smoothing is specified for zero probabilities, so the loss can be undefined unless an unstated safeguard is added.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semi-Supervised Multi-Task Learning With Chest X-Ray Images." pith.science (2026). https://pith.science/paper/7KXIB2IZ

@misc{pith2026190803693,
  author       = {Pith},
  title        = {Pith review of: Semi-Supervised Multi-Task Learning With Chest X-Ray Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7KXIB2IZ}},
  note         = {Machine review of arXiv:1908.03693}
}
read the original abstract

Discriminative models that require full supervision are inefficacious in the medical imaging domain when large labeled datasets are unavailable. By contrast, generative modeling---i.e., learning data generation and classification---facilitates semi-supervised training with limited labeled data. Moreover, generative modeling can be advantageous in accomplishing multiple objectives for better generalization. We propose a novel multi-task learning model for jointly learning a classifier and a segmentor, from chest X-ray images, through semi-supervised learning. In addition, we propose a new loss function that combines absolute KL divergence with Tversky loss (KLTV) to yield faster convergence and better segmentation performance. Based on our experimental results using a novel segmentation model, an Adversarial Pyramid Progressive Attention U-Net (APPAU-Net), we hypothesize that KLTV can be more effective for generalizing multi-tasking models while being competitive in segmentation-only tasks.

Figures

Figures reproduced from arXiv: 1908.03693 by the authors.

Figure 1
Figure 1. (a) Basic structure of the proposed APPAU-Net model. The segmentor [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the segmentor or PPAU-Net in our APPAU-Net model. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visual comparison of the lung segmentation by the APPAU-Net model [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual comparison of the lung segmentation in an abnormal (TB) X-Ray [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 19 canonical work pages

  1. [1]

    In: International Conf

    Caruana, R.: Multitask learning: A knowledge-based source of inductive bias. In: International Conf. on Machine Learning. (1993)

    work page 1993

  2. [2]

    WPC (2018)

    Mittal, A., Hooda, R., Sofat, S.: Lf-segnet: A fully convolutional encoder–decoder network for segmenting lung fields from chest radiographs. WPC (2018)

    work page 2018

  3. [3]

    In: ISBI

    Zhu, W., Xiang, X., Tran, T.D., Hager, G.D., Xie, X.: Adversarial deep structured nets for mass segmentation from mammograms. In: ISBI. (2018)

    work page 2018

  4. [4]

    Semantic-Aware Generative Adversarial Nets for Unsupervised Domain Adaptation in Chest X-ray Segmentation

    Chen, C., Dou, Q., et al.: Semantic-aware generative adversarial nets for unsuper- vised domain adaptation in chest x-ray segment. arXiv:1806.00600 (2018)

    work page Pith review arXiv 2018

  5. [5]

    In: Deep Learning in Medical Image Analysis

    Dai, W., Dong, N., Wang, Z., Liang, X., Zhang, H., Xing, E.P.: Scan: Structure correcting adversarial network for organ segmentation in chest x-rays. In: Deep Learning in Medical Image Analysis. Volume 11045 of LNCS. (2018)

    work page 2018

  6. [6]

    In: Proc

    Madani, A., et al.: Semi-supervised learning with GANs for chest X-ray classification with ability of data domain adaptation. In: Proc. ISBI. (2018)

    work page 2018

  7. [7]

    arXiv:1801.09927 (2018)

    Guan, Q., Huang, Y., et al.: Diagnose like a radiologist: Attention guided convolu- tional net for thorax disease classification. arXiv:1801.09927 (2018)

    arXiv 2018

  8. [8]

    Multi-Adversarial Variational Autoencoder Networks

    Imran, A.A.Z., Terzopoulos, D.: Multi-adversarial variational autoencoder networks. arXiv preprint arXiv:1906.06430 (2019)

    work page Pith review arXiv 2019

Show all 22 references
  1. [9]

    CoRR (2018)

    Rezaei, M., Yang, H., et al.: Multi-task generative adversarial network for handling imbalanced clinical data. CoRR (2018)

    2018

  2. [10]

    AI in Med (2019)

    Girard, F., Kavalec, C., Cheriet, F.: Joint segmentation and classification of retinal arteries/veins from fundus images. AI in Med (2019)

    2019

  3. [11]

    In: MICCAI

    Mehta, S., Mercan, E., Bartlett, J., et al.: Y-Net: Joint segmentation and classifica- tion for diagnosis of breast biopsy images. In: MICCAI. (2018)

    2018

  4. [12]

    arXiv:1703.01025 (2017)

    Yang, X., Zeng, Z., Yeo, S.Y., et al.: A novel multi-task deep learning model for skin lesion segmentation and classification. arXiv:1703.01025 (2017)

    2017 arXiv

  5. [13]

    In: EMBC

    Khosravan, N., Bagci, U.: Semi-supervised multi-task learning for lung cancer diagnosis. In: EMBC. (2018)

    2018

  6. [14]

    IEEE TMI (2018)

    Fu, H., Cheng, J., Xu, Y., et al.: Joint optic disc and cup segmentation based on multi-label deep network and polar transformation. IEEE TMI (2018)

    2018

  7. [15]

    In: DLMIA

    Imran, A.A.Z., Hatamizadeh, A., Ananth, S.P., et al.: Automatic segmentation of pulmonary lobes using a progressive dense V-network. In: DLMIA. (2018)

    2018

  8. [16]

    arXiv:1804.03999 (2018)

    Oktay, O., Schlemper, J., Folgoc, L.L., et al.: Attention U-net: Learning where to look for the pancreas. arXiv:1804.03999 (2018)

    2018 arXiv

  9. [17]

    Quant Imag in Med and Surg (2014)

    Jaeger, S., Candemir, S., et al.: Two public chest X-ray datasets for computer-aided screening of pulmonary diseases. Quant Imag in Med and Surg (2014)

    2014

  10. [18]

    J of Roent (2000) Semi-Supervised Multi-Task Learning With Chest X-Ray Images 9 Table 3: Partitioning of the image datasets

    Shiraishi, J., Katsuragawa, S., et al.: Development of a digital image database for chest radiographs with and without a lung nodule. J of Roent (2000) Semi-Supervised Multi-Task Learning With Chest X-Ray Images 9 Table 3: Partitioning of the image datasets. Dataset Train Vali...

    2000

  11. [19]

    This dataset contains separate left and right lung ground truth masks, which we combined in our experiments

    MCX: In this dataset, there are 138 frontal X-Rays: 80 X-Rays are normal and 58 X-Rays show manifestations of Tuberculosis. This dataset contains separate left and right lung ground truth masks, which we combined in our experiments

  12. [20]

    Of them, 336 are normal X-Rays and 326 are abnormal cases with manifestations of TB, including pediatric X-rays

    SCX: This dataset comprises 662 frontal chest X-rays. Of them, 336 are normal X-Rays and 326 are abnormal cases with manifestations of TB, including pediatric X-rays. After carefully examining all the cases, we selected 527 X-rays in good agreement with the corresponding groun...

  13. [21]

    In addition to the lung masks (separated left-right), this dataset includes ground truth masks for the heart and clavicles (separated left-right)

    JSRT: This dataset contains 247 chest X-rays in which 154 images show pulmonary lung nodules and 93 images show no lung nodules. In addition to the lung masks (separated left-right), this dataset includes ground truth masks for the heart and clavicles (separated left-right)

  14. [22]

    We split it into three sets: training set (615), validation set (69), and testing set (228)

    CCX: Combining the above three datasets, we created a dataset of 912 X-ray images, that we dubbed the CCX (Combined chest X-ray) dataset. We split it into three sets: training set (615), validation set (69), and testing set (228). The models were trained on the training set an...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.