SegX: Improving Interpretability of Clinical Image Diagnosis with Segmentation-based Enhancement

Mingcheng Zhu; Yuhao Zhang; Zhiyao Luo

REVIEW 4 major objections 5 minor 29 references

SegX: Improving Interpretability of Clinical Image Diagnosis with Segmentation-based Enhancement

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SegX aligns AI explanation maps with clinically relevant regions and uses that alignment to score prediction certainty.

desk verdict A cleanly written extension of Yuan et al.'s mask-overlay idea to new datasets, but the headline result is an arithmetic identity, not an empirical finding. read the letter →

arxiv 2502.10296 v1 pith:CKCEM446 submitted 2025-02-14 eess.IV

Yuhao Zhang , Mingcheng Zhu , Zhiyao Luo This is my paper

classification eess.IV

keywords explainableAImedicalimageinterpretationsegmentation-guidedexplanationuncertaintyestimationGrad-CAMSHAPdermoscopychestX-ray

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

SegX is a plug-and-play method for making existing XAI heatmaps in medical imaging point at clinically relevant anatomy instead of background artifacts. It takes the segmentation mask of the disease region and intersects it with the explanation map produced by any XAI method, yielding a refined map. The companion SegU score measures how much the original explanation overlaps the segmentation mask and uses that overlap as a certainty estimate for the model's prediction. On dermoscopic and chest X-ray datasets, the paper reports that SegX consistently raises IoU and AUITC alignment scores across ResNet and DenseNet with Grad-CAM and SHAP, and that SegU separates correct from incorrect predictions in most settings. The point is to give clinicians explanations that match what they look at, while flagging cases where the model's reasoning is off-target.

What carries the argument

The load-bearing object is the segmentation-guided mask $m^{\mathrm{segX}}_{i,j} = m^{\mathrm{seg}}_{i,j} \cdot m^{\mathrm{xai}}_{i,j}$, where $m^{\mathrm{seg}}$ is the UNet mask of the clinically relevant region and $m^{\mathrm{xai}}$ is the saliency map from any XAI method; the refined map keeps only the intersection of model focus and clinical interest. The certainty side rests on $c_{\mathrm{IoU}} = |m^{\mathrm{xai}} \cap m^{\mathrm{seg}}|/|m^{\mathrm{xai}} \cup m^{\mathrm{seg}}|$ and its threshold-averaged version $c_{\mathrm{AUITC}} = \int_0^1 c_{\mathrm{IoU}}(\tau)\, d\tau$. The UNet is trained with a composite cross-entropy plus Dice loss, and the same masks serve both to refine explanations and to score their reliability.

What would settle it

Run the same SegX/SegU pipeline with a deliberately wrong segmentation mask, such as a randomly shifted or background-region mask; if the IoU and AUITC gains over baseline XAI persist, the reported improvements are an artifact of mask intersection rather than clinical relevance. A blinded clinician study that finds original explanations as useful as SegX maps would also falsify the interpretability claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a clinical-region segmentation mask acts as a reliable guide for post-hoc explanations: rather than changing the classifier or the XAI method, SegX overlays the segmentation mask on the explanation map, so the final interpretation only shows the model's attention inside clinically significant areas. The paper further claims that the degree of alignment between an unrefined explanation and the segmentation mask is itself a useful certainty signal: the IoU-based score $c_{\mathrm{IoU}}$ and the threshold-integrated score $c_{\mathrm{AUITC}}$ are consistently higher for correct predictions than for incorrect ones across the tested datasets, models, and XAI methods. The claim is stronger than 'explanations look better': it asserts that this alignment-based score reliably reflects whether the classifier's prediction is trustworthy, giving clinicians a numeric flag for which predictions need extra scrutiny.

Load-bearing premise

The load-bearing premise is that the UNet's segmentation mask correctly identifies the clinically relevant region, so that overlap with that mask is a valid measure of explanation quality and prediction certainty.

Editorial extensions

If this is right

SegX can be dropped onto any trained classifier and any XAI method as a post-hoc filter, so it does not require retraining the diagnostic model.
SegU gives clinicians a threshold-free certainty score that can be read alongside a prediction, flagging low-alignment cases for manual review.
Because the gains depend on baseline XAI quality, SegX is most useful where the original heatmap already has some signal; on the chest X-ray set the reported improvements are under one percentage point.
The approach works in a multi-label setting, where each predicted disease can be paired with its own clinical-region mask.

Reading between the lines

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

Because the reported evaluation measures overlap with the very same segmentation masks SegX uses to filter, part of the IoU improvement is arithmetic; a stronger test would compare against independent expert annotations or clinician rankings.
SegU's failure to separate correct from incorrect SHAP explanations on dermoscopic images suggests the alignment metric inherits the XAI method's assumptions; dispersed attributions can score similarly whether or not the prediction is right.
SegX could be paired with zero-shot segmentation models, which the paper notes as future direction, removing the need for per-disease segmentation training.
One testable extension is to compare SegU's certainty score against the classifier's own softmax confidence: if SegU adds predictive accuracy beyond confidence, it earns a place as a separate clinical signal.

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.

Desk Editor's Note

read the letter

Quick take: this is a well-organized, readable paper with a public codebase and honest citations of prior work. The genuinely new bits are modest: applying the segmentation-mask overlay to HAM10000 and ChestX-Det10, across two classifiers and two XAI methods, plus the SegU idea of using overlap with the clinical mask as a certainty score. The authors also deserve credit for stating their limitations clearly, including the caveat that IoU may not be the right metric.

The problem is that the central quantitative claim is not supported by the experiment. SegX, as defined in Eq. 7, intersects the XAI map with the segmentation mask. The evaluation in Table 3 then measures IoU with that same segmentation mask. It is a mathematical identity that masking an explanation by m_seg cannot decrease IoU with m_seg, and it strictly increases whenever the original explanation has any mass outside the mask. So the "consistent improvement" in Table 3 is forced by construction; it says nothing about interpretability. The paper would need evaluation with human raters, or at least a metric not built from the same mask, to support its claim.

The SegU results are also weaker than the abstract suggests. Table 4 has no error bars or significance tests, and several rows show near-identical scores between correct and incorrect groups. The paper itself acknowledges this for SHAP on HAM10000, but the conclusion still says SegU "reliably reflects the correctness" of predictions. That overstates what the data show.

There is also an internal tension: the introduction criticizes existing KG-XAI methods for masking out background regions, and then SegX does exactly that. The addition of SegU partially addresses this, but SegU uses the same segmentation mask, so it does not resolve the concern.

My advice: if this lands on your desk, do not reject it out of hand, but do send it to a referee who can pinpoint the circular evaluation. The paper is short and the flaw is fixable in revision. As it stands, however, the headline claim is an artifact of the metric, and the uncertainty claim is not backed by statistical evidence. I would not cite it as a demonstration of improved interpretability, but it could be a starting point for a properly controlled study. My call: peer review, but with the expectation of major revisions.

Referee Report

4 major / 5 minor

Summary. The paper proposes SegX, a post-hoc module that intersects XAI saliency maps with a UNet segmentation mask of clinically relevant regions, and SegU, an uncertainty score based on IoU between a saliency map and the segmentation mask. Experiments on HAM10000 and ChestX-Det10 with ResNet/DenseNet and SHAP/Grad-CAM report that SegX improves IoU and AUITC, and that SegU separates correct from incorrect predictions. The manuscript claims that both modules are model-agnostic and improve the interpretability and reliability of medical image diagnosis.

Significance. If the central claims were supported, SegX would be a simple and practical post-processing step for medical XAI. However, the quantitative evidence is not independent of the method itself: SegX is the intersection of the XAI map with a segmentation mask, and the main evaluation metric measures overlap with the same kind of mask. No human-grounded evaluation, no comparison with prior knowledge-guided XAI baselines, and no statistical characterization of the SegU scores are provided. The manuscript does ship code and applies the approach to two imaging modalities, which are useful starting points, but the current evaluation does not establish improved interpretability or reliable uncertainty assessment.

major comments (4)

[Segmentation-based Explanation (SegX); Eq. (7); Table 3] The central quantitative claim in Table 3 is an artifact of the metric if the reference mask is m_seg. SegX is defined in Eq. (7) as m_segX = m_seg * m_xai, so all explanation mass outside m_seg is discarded. When IoU is computed against m_seg, IoU(m_segX, m_seg) = |m_xai ∩ m_seg| / |m_seg|, whereas IoU(m_xai, m_seg) = |m_xai ∩ m_seg| / |m_xai ∪ m_seg|, and the former is always greater than or equal to the latter because |m_seg| ≤ |m_xai ∪ m_seg|. Thus every row of Table 3 must improve even for a random or adversarial XAI map. If the evaluation instead uses an independent clinical ground-truth mask, the manuscript never states this, never reports the segmentation model's test accuracy, and never explains where such masks come from for datasets that do not provide lesion-level segmentation labels. In either reading, Table 3 cannot establish improved interpretability.
[Experiment Settings; Eq. (8)] Segmentation quality is unvalidated. Both SegX and SegU rely on m_seg approximating clinically relevant regions, and the segmentation model is trained with the composite loss in Eq. (8), yet the paper reports no Dice or IoU of the UNet on the test sets, no source of the segmentation labels used for HAM10000 (which has no official lesion segmentation masks), and no description of how the ChestX-Det10 bounding-box annotations were converted to masks. Without this information, the reader cannot assess whether m_seg is a clinically meaningful reference or a learned artifact.
[Segmentation-based Uncertainty Assessment (SegU); Table 4] The claim that SegU 'reliably reflects the correctness of the model's predictions' is not supported by the reported statistics. Table 4 gives only point estimates with no confidence intervals, hypothesis tests, AUC, or calibration analysis. For example, on HAM10000 with ResNet and SHAP, the cIoU values are 0.133 for correct and 0.122 for incorrect predictions, and with DenseNet they are 0.124 and 0.123. The paper even acknowledges in the subsection 'SegU Struggles with SHAP on Dermoscopic Images' that these gaps are marginal. Such tiny differences cannot sustain a general reliability claim.
[Related Work; Conclusion] Interpretability is never validated against human judgment or against the prior knowledge-guided XAI methods discussed in the Related Work. The paper contrasts SegX with the segmentation-based approach of Yuan et al. (2023) and the template-based method of Yuan et al. (2024), but Table 3 contains no comparison with either method. The central interpretability claim rests entirely on IoU with the same type of mask used to construct SegX, and the limitation section itself concedes that 'IoU may not be the best measurement to calculate the distance between clinically relevant regions and XAI outputs.' A human-grounded evaluation or an independent clinical-outcome measure is needed to support the conclusion.

minor comments (5)

[Abstract] The abstract says 'across mortalities,' which appears to be a typo for 'across modalities.'
[Segmentation-based Uncertainty Assessment (SegU); Eq. (9)] The text preceding Eq. (9) says the score quantifies overlap between the XAI map and 'the corresponding segmentation mask mxai_i,j,' but the symbol used in the equation is mseg_i,j; the prose should refer to mseg_i,j.
[Segmentation-based Uncertainty Assessment (SegU); Eq. (10)] Equation (10) defines cAUITC as the integral of cIoU(τ), but cIoU in Eq. (9) is defined without a threshold argument; the paper should define how the threshold τ produces a binary map mxai(τ) and should make the dependence of cIoU on τ explicit.
[Results of SegX's Effectiveness Evaluation; Table 3] The table title 'IoU@5%' is not defined in the text; the paper should explain how the top-5% saliency threshold is applied to compute the IoU value.
[Experiment; Datasets] The paper should clarify the origin of the clinical-region masks used for training and evaluation, since HAM10000 does not provide official lesion segmentation masks and ChestX-Det10 is a detection dataset with bounding boxes rather than segmentation masks.

Circularity Check

1 steps flagged · score 8.0 of 10

SegX's 'consistent improvement' is an algebraic consequence of masking the explanation by the same mask used as the evaluation reference.

self definitional [Methodology Eq. (7); Results 'Results of SegX's Effectiveness Evaluation' Eqs. (11)-(12) and Table 3]
"msegX_{i,j} = mseg_{i,j} · mxai_{i,j} (7) ... The AUITC metric provides a comprehensive assessment of alignment by evaluating IoU scores across varying thresholds τ applied to XAI outputs. It is computed as: AUITC = ∫_0^1 IoU(mgt, m(τ)) dτ, where mgt is the ground truth clinically significant mask (12)."

Under the paper's operationalization, the reference mask mgt is the same clinically significant mask as mseg: mseg is defined as the mask of clinically relevant regions, and no independent clinical ground truth is introduced for Table 3. Since msegX = mxai ∩ mseg ⊆ mseg, IoU(msegX, mseg) = |mxai∩mseg| / |mseg| ≥ |mxai∩mseg| / |mxai∪mseg| = IoU(mxai, mseg), with strict increase whenever mxai has any positive mass outside mseg. Therefore every row of Table 3 must show SegX ≥ Original; the claimed 'consistent improvement' is an algebraic identity, not an empirical finding about interpretability. AUITC integrates the same pointwise inequality.

full rationale

The central quantitative claim of the paper is that SegX improves interpretability, measured by IoU and AUITC. However, SegX is defined in Eq. (7) as the intersection of the XAI map with the segmentation mask mseg. The evaluation in Eqs. (11)-(12) computes IoU and AUITC against the same clinically significant mask mgt that mseg is supposed to represent. For any XAI map, masking by mseg necessarily increases IoU with mseg, because the denominator shrinks from |mxai ∪ mseg| to |mseg| while the numerator stays |mxai ∩ mseg|. Thus Table 3's consistent improvement is a mathematical consequence of the method's definition, not an empirical discovery. Similarly, SegU's certainty scores (Eqs. (9)-(10)) are derived from the same overlap and are used to claim reliable reflection of prediction correctness; the paper provides no calibration, AUC, significance testing, or independent validation, and its own limitation section admits IoU may be inappropriate. This makes the central claims circular by construction, warranting a score of 8. The paper does not rely on self-citation chains or imported uniqueness theorems; the circularity is self-definitional.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method relies on two external components: a trained segmentation model and a pretrained classifier. The free parameters are the saliency threshold and the segmentation loss weight. The key axioms are domain assumptions about the validity of segmentation masks as clinical ground truth and about the relationship between attention and correctness. No new physical or conceptual entities are introduced.

free parameters (2)

Saliency threshold for explanation maps = 5%
The paper extracts the top 5% of the XAI output as the explanation map, following prior work. This choice affects all IoU and AUITC values and is a hand-picked threshold.
Segmentation loss weight lambda = Not reported
Eq. (8) combines cross-entropy and Dice loss with a weight lambda, but the value is not given. The segmentation quality depends on this hyperparameter, and the paper does not state how it was chosen.

assumptions (3)

domain assumption Segmentation masks represent the clinically relevant region.
SegX and SegU both rely on the UNet output as a proxy for clinical interest, but the paper does not validate the segmentation accuracy on the test sets or assess how segmentation errors propagate to the explanation refinement.
domain assumption Correct predictions are more likely to be made from clinically significant regions.
This hypothesis is stated in the SegU section but is not derived, tested, or compared with alternative uncertainty sources.
domain assumption IoU between the explanation map and the segmentation mask measures interpretability.
The evaluation equates interpretability with overlap with clinical masks. No human evaluation, clinical utility test, or alternative interpretability metric is used.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SegX: Improving Interpretability of Clinical Image Diagnosis with Segmentation-based Enhancement." pith.science (2026). https://pith.science/paper/CKCEM446

@misc{pith2026250210296,
  author       = {Pith},
  title        = {Pith review of: SegX: Improving Interpretability of Clinical Image Diagnosis with Segmentation-based Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CKCEM446}},
  note         = {Machine review of arXiv:2502.10296}
}

read the original abstract

Deep learning-based medical image analysis faces a significant barrier due to the lack of interpretability. Conventional explainable AI (XAI) techniques, such as Grad-CAM and SHAP, often highlight regions outside clinical interests. To address this issue, we propose Segmentation-based Explanation (SegX), a plug-and-play approach that enhances interpretability by aligning the model's explanation map with clinically relevant areas leveraging the power of segmentation models. Furthermore, we introduce Segmentation-based Uncertainty Assessment (SegU), a method to quantify the uncertainty of the prediction model by measuring the 'distance' between interpretation maps and clinically significant regions. Our experiments on dermoscopic and chest X-ray datasets show that SegX improves interpretability consistently across mortalities, and the certainty score provided by SegU reliably reflects the correctness of the model's predictions. Our approach offers a model-agnostic enhancement to medical image diagnosis towards reliable and interpretable AI in clinical decision-making.

Figures

Figures reproduced from arXiv: 2502.10296 by the authors.

**Figure 1.** Negative examples of using XAI to highlight highly important regions. The first column shows the original images from HAM10000 and ChestX-Det10, respectively. The second column shows the clinical interests of each image. The last column shows the Grad-CAM interpreted results on a learned model. It is clear to see that the model overemphasises background area for making a diagnosis, which mismatches clinical inte… view at source ↗

**Figure 2.** Specifically, SegX applies the segmentation masks of [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 2.** Overview of the proposed method for improving XAI outputs with segmentation-based improvement. The process begins with the classification and XAI pipeline (green), where a medical image is processed by a classification model to generate predictions and an XAI method produces the original explanation map. Afterwards, the segmentation-based enhancement (blue) starts with a segmentation model that generates a clinica… view at source ↗

Figures from the paper (1 more)

**Figure 3.** Figure 3: Examples of original XAI masks (baseline) versus SegX masks on the HAM10000 and ChestX-Det10 test sets. The model’s top 5% focus region is highlighted in white for the baseline and in red for the SegX method. Experiment Datasets To evaluate the effectiveness of our pro…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Aas, K.; Jullum, M.; and L land, A. 2021. Explaining individual predictions when features are dependent: More accurate approximations to Shapley values. Artificial Intelligence, 298: 103502

2021
[4]

J.; Salvi, M.; Molinari, F.; and Acharya, U

Abbasian Ardakani, A.; Airom, O.; Khorshidi, H.; Bureau, N. J.; Salvi, M.; Molinari, F.; and Acharya, U. R. 2024. Interpretation of Artificial Intelligence Models in Healthcare: A Pictorial Guide for Clinicians. Journal of Ultrasound in Medicine, 43(10): 1789--1818

work page 2024
[5]

C.; Tschandl, P.; Rinner, C.; Akay, B

Barata, C.; Rotemberg, V.; Codella, N. C.; Tschandl, P.; Rinner, C.; Akay, B. N.; Apalla, Z.; Argenziano, G.; Halpern, A.; Lallas, A.; et al. 2023. A reinforcement learning model for AI-based decision support in skin cancer. Nature Medicine, 29(8): 1941--1946

work page 2023
[6]

A.; Nauta, M.; Seifert, C.; Kr \"a mer, N.; Friedrich, C

Borys, K.; Schmitt, Y. A.; Nauta, M.; Seifert, C.; Kr \"a mer, N.; Friedrich, C. M.; and Nensa, F. 2023. Explainable AI in medical imaging: An overview for clinical practitioners--Beyond saliency-based XAI approaches. European Journal of Radiology, 162: 110786

work page 2023
[7]

Cozzi, D.; Cavigli, E.; Moroni, C.; Smorchkova, O.; Zantonelli, G.; Pradella, S.; and Miele, V. 2021. Ground-glass opacity (GGO): a review of the differential diagnosis in the era of COVID-19. Japanese Journal of Radiology, 39(8): 721--732

work page 2021
[8]

Dai, T.; Zhang, R.; Hong, F.; Yao, J.; Zhang, Y.; and Wang, Y. 2024. UniChest: Conquer-and-Divide Pre-training for Multi-Source Chest X-Ray Classification. IEEE Transactions on Medical Imaging

work page 2024

Show all 29 references

[9]

Gu, J.; Wang, Z.; Kuen, J.; Ma, L.; Shahroudy, A.; Shuai, B.; Liu, T.; Wang, X.; Wang, G.; Cai, J.; et al. 2018. Recent advances in convolutional neural networks. Pattern Recognition, 77: 354--377

2018
[10]

Hassija, V.; Chamola, V.; Mahapatra, A.; Singal, A.; Goel, D.; Huang, K.; Scardapane, S.; Spinelli, I.; Mahmud, M.; and Hussain, A. 2024. Interpreting black-box models: a review on explainable artificial intelligence. Cognitive Computation, 16(1): 45--74

2024
[11]

He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770--778

2016
[12]

Huang, G.; Liu, Z.; Van Der Maaten, L.; and Weinberger, K. Q. 2017. Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 4700--4708

2017
[13]

Jung, H.-G.; Nam, W.-J.; Kim, H.-W.; and Lee, S.-W. 2023. Weakly supervised thoracic disease localization via disease masks. Neurocomputing, 517: 34--43

2023
[14]

Kato, S.; and Hotta, K. 2024. Adaptive t-vMF dice loss: An effective expansion of dice loss for medical image segmentation. Computers in Biology and Medicine, 168: 107695

2024
[15]

Liu, J.; Lian, J.; and Yu, Y. 2020. ChestX-Det10: chest x-ray dataset on detection of thoracic abnormalities. arXiv preprint arXiv:2006.10550

2020 arXiv
[16]

W.; Ooi, C

Loh, H. W.; Ooi, C. P.; Seoni, S.; Barua, P. D.; Molinari, F.; and Acharya, U. R. 2022. Application of explainable artificial intelligence for healthcare: A systematic review of the last decade (2011--2022). Computer Methods and Programs in Biomedicine, 226: 107161

2022
[17]

M.; and Lee, S.-I

Lundberg, S. M.; and Lee, S.-I. 2017. A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems, 30

2017
[18]

K.; Singh, P

Mall, P. K.; Singh, P. K.; Srivastav, S.; Narayan, V.; Paprzycki, M.; Jaworska, T.; and Ganzha, M. 2023. A comprehensive review of deep neural networks for medical image processing: Recent developments and future opportunities. Healthcare Analytics, 100216

2023
[19]

Pinto-Coelho, L. 2023. How artificial intelligence is shaping medical imaging technology: A survey of innovations and applications. Bioengineering, 10(12): 1435

2023
[20]

K.; Swapna, K.; Datta, D.; and Rajest, S

Ramesh, K.; Kumar, G. K.; Swapna, K.; Datta, D.; and Rajest, S. S. 2021. A review of medical image segmentation algorithms. EAI Endorsed Transactions on Pervasive Health and Technology, 7(27): e6--e6

2021
[21]

Ronneberger, O.; Fischer, P.; and Brox, T. 2015. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, 234--241. Springer

2015
[22]

R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; and Batra, D

Selvaraju, R. R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; and Batra, D. 2017. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision, 618--626

2017
[23]

Tschandl, P.; Rosendahl, C.; and Kittler, H. 2018. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data, 5(1): 1--9

2018
[24]

Yeung, M.; Sala, E.; Sch \"o nlieb, C.-B.; and Rundo, L. 2022. Unified focal loss: Generalising dice and cross entropy-based losses to handle class imbalanced medical image segmentation. Computerized Medical Imaging and Graphics, 95: 102026

2022
[25]

Yuan, H.; Hong, C.; Jiang, P.-T.; Zhao, G.; Tran, N. T. A.; Xu, X.; Yan, Y. Y.; and Liu, N. 2024. Clinical domain knowledge-derived template improves post hoc AI explanations in pneumothorax classification. Journal of Biomedical Informatics, 104673

2024
[26]

Yuan, H.; Hong, C.; Tran, N. T. A.; Xu, X.; and Liu, N. 2023. Leveraging Anatomical Constraints with Uncertainty for Pneumothorax Segmentation. arXiv preprint arXiv:2311.15213

2023 arXiv
[27]

Zhang, X.; Qian, B.; Cao, S.; Li, Y.; Chen, H.; Zheng, Y.; and Davidson, I. 2020. INPREM: An interpretable and trustworthy predictive model for healthcare. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 450--460

2020
[28]

Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; and Torralba, A. 2016. Learning deep features for discriminative localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2921--2929

2016
[29]

Zhu, M.; Zhang, R.; and Wang, H. 2022. Recognizing irrelevant faces in short-form videos based on feature fusion and active learning. Neurocomputing, 501: 694--704

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Aas, K.; Jullum, M.; and L land, A. 2021. Explaining individual predictions when features are dependent: More accurate approximations to Shapley values. Artificial Intelligence, 298: 103502

2021

[4] [4]

J.; Salvi, M.; Molinari, F.; and Acharya, U

Abbasian Ardakani, A.; Airom, O.; Khorshidi, H.; Bureau, N. J.; Salvi, M.; Molinari, F.; and Acharya, U. R. 2024. Interpretation of Artificial Intelligence Models in Healthcare: A Pictorial Guide for Clinicians. Journal of Ultrasound in Medicine, 43(10): 1789--1818

work page 2024

[5] [5]

C.; Tschandl, P.; Rinner, C.; Akay, B

Barata, C.; Rotemberg, V.; Codella, N. C.; Tschandl, P.; Rinner, C.; Akay, B. N.; Apalla, Z.; Argenziano, G.; Halpern, A.; Lallas, A.; et al. 2023. A reinforcement learning model for AI-based decision support in skin cancer. Nature Medicine, 29(8): 1941--1946

work page 2023

[6] [6]

A.; Nauta, M.; Seifert, C.; Kr \"a mer, N.; Friedrich, C

Borys, K.; Schmitt, Y. A.; Nauta, M.; Seifert, C.; Kr \"a mer, N.; Friedrich, C. M.; and Nensa, F. 2023. Explainable AI in medical imaging: An overview for clinical practitioners--Beyond saliency-based XAI approaches. European Journal of Radiology, 162: 110786

work page 2023

[7] [7]

Cozzi, D.; Cavigli, E.; Moroni, C.; Smorchkova, O.; Zantonelli, G.; Pradella, S.; and Miele, V. 2021. Ground-glass opacity (GGO): a review of the differential diagnosis in the era of COVID-19. Japanese Journal of Radiology, 39(8): 721--732

work page 2021

[8] [8]

Dai, T.; Zhang, R.; Hong, F.; Yao, J.; Zhang, Y.; and Wang, Y. 2024. UniChest: Conquer-and-Divide Pre-training for Multi-Source Chest X-Ray Classification. IEEE Transactions on Medical Imaging

work page 2024