Pith. sign in

REVIEW 4 major objections 7 minor 60 references

Interpreting Radiologist's Intention from Eye Movements in Chest X-ray Diagnosis

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read RadGazeIntent reads which chest X-ray finding a radiologist is examining from eye fixations.

desk verdict A novel task and solid engineering, but the ground-truth labels make 'intention' a function of report timing rather than gaze position, so the central claim outruns the evidence. read the letter →

arxiv 2507.12461 v1 pith:WIQINTAL submitted 2025-07-16 cs.CV cs.AI

Trong-Thang Pham , Anh Nguyen , Zhigang Deng , Carol C. Wu , Hien Van Nguyen , Ngan Le This is my paper · ORCID
classification cs.CVcs.AI
keywords eyegazechestX-rayradiologistintentionfixationclassificationtransformervisualsearchmedicalimageanalysismulti-label
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

This paper argues that the diagnostic purpose behind each radiologist eye fixation can be recovered from gaze data, and presents RadGazeIntent to do it: a transformer that takes a chest X-ray and a sequence of fixations and labels each fixation with the finding the radiologist is examining at that moment. Because intention cannot be observed directly, the authors construct three benchmark datasets from two existing eye-tracking corpora, each encoding a different assumption about how radiologists search: systematic checklist-driven search, uncertainty-driven exploration, and a hybrid of broad scanning followed by focused search. The model combines fine-grained foveal features at the fixation point with coarse peripheral context under a causal attention mask that lets earlier fixations inform later ones but not vice versa. The reported result is that RadGazeIntent predicts fixation-level intention better than MLP, LSTM, heatmap-plus-RNN, and scanpath baselines across all three settings. A sympathetic reader would care because interpreting what a radiologist is looking for, rather than merely where they look, could enable gaze-guided reporting, training feedback, and assistive systems that respond to diagnostic focus.

What carries the argument

The load-bearing mechanism is the Peripheral-aware Causal Self-Attention block, whose mask gives every fixation token access to all peripheral image tokens and to earlier fixations, while blocking future fixations. Before that, the Structural Embedding fovea-maps each fixation coordinate into the high-resolution feature map, adds 2D spatial and 1D temporal embeddings, and concatenates those tokens with flattened low-resolution peripheral tokens; the Intention Decoder then applies Pool Attention to compress the fixation sequence into few tokens and cross-attention to decode each fixation's finding confidence. The label-construction rules in the dataset section are equally load-bearing: RadExplore labels a fixation by whether its timestamp precedes a dictation sentence mentioning a finding, RadSeq restricts this to the time window of a finding's search, and RadHybrid unions RadSeq with the first second of scanning.

What would settle it

Use a think-aloud or retrospective-recall protocol: ask radiologists at sampled fixations what they were looking for, and compare their stated intentions against the temporal-containment labels computed from their own dictation. If stated intention and those labels disagree on fixations that precede the first mention of a finding, or if a fixated but never-dictated finding is assigned label zero, the central claim is falsified.

Watch

Extended reading notes

Core claim

RadGazeIntent's central claim is that each radiologist fixation on a chest X-ray carries a recoverable diagnostic intention, and that intention can be decoded from the fixation sequence plus the image itself. The paper frames intention as a multi-label sequence labeling problem: given fixations with coordinates and durations, the model outputs confidence scores over radiological findings for each fixation. The authors test the claim under three definitions of intention, derived from two public eye-tracking corpora, and report that RadGazeIntent outperforms MLP, LSTM, a heatmap-based temporal classifier, and ChestSearch on accuracy, F1, precision, and recall across all three settings on both corpora. In the authors' telling, the result shows that a transformer with peripheral context and causal temporal masking can turn noisy, fine-grained fixation data into stable representations of what the radiologist is looking for, not merely where they are looking.

Load-bearing premise

The load-bearing premise is that a fixation's intention can be read off the dictated report: a fixation counts as intending a finding whenever that finding is mentioned in any later sentence, so if temporal containment in dictation is a poor proxy for what the radiologist was actually thinking, the model's labels and therefore its reported success lose their meaning.

Editorial extensions

If this is right

  • A reliable per-fixation intention signal enables real-time interactive reading: a system could know which finding a radiologist is actively checking and offer context-specific assistance or draft the corresponding report sentence.
  • The three datasets let the field evaluate gaze-AI under competing theories of radiologist search behavior rather than committing to a single definition of intention.
  • Because predictions carry confidence scores per finding, low-confidence moments can be surfaced as opportunities for a second look or for asking the radiologist to verify a region.
  • Ablations show the model's signal is carried jointly by foveal features at the fixation point, peripheral context, temporal order, and spatial position, meaning each of these channels contributes to decoding intention.

Reading between the lines

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

  • Because the label rule requires a finding to appear somewhere in the dictated report, the benchmark can never credit a radiologist for fixating a finding they then fail to report; measuring missed findings would need a different label source, such as image-level ground-truth abnormalities.
  • The causal mask enforces that earlier fixations cannot use later report evidence; this makes the model's per-fixation predictions a plausible candidate for online use during the actual scan, a property the paper notes but does not turn into a real-time experiment.
  • A natural extension is to test the same architecture on other modalities with report-like ground truth, such as CT or pathology, since the paper's own limitation section predicts behavior may differ there.
  • The authors set ChestSearch's radius filter to zero to keep all fixations; an ablation varying the radius could determine how much of the performance gap is due to retaining temporally sparse fixations rather than architectural superiority.
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 proposes RadGazeIntent, a transformer-based model that classifies each eye fixation of a radiologist reading a chest X-ray according to the diagnostic finding the radiologist is supposedly examining. The authors create three intention-labeled datasets, RadSeq, RadExplore, and RadHybrid, by post-processing the public EGD and REFLACX gaze datasets under three behavioral assumptions. The model combines foveal and peripheral image features, structural spatial/temporal embeddings, and causal self-attention, and is evaluated against MLP, LSTM, Karargyris et al., and ChestSearch baselines with reported F1 scores around 70%. The central claim is that RadGazeIntent predicts which findings radiologists are examining at specific moments and outperforms all baselines.

Significance. The paper targets a real gap: existing gaze-assisted chest X-ray work predicts saliency or scanpaths rather than explaining the purpose of individual fixations. The proposed architecture is thoughtful, with foveal/peripheral feature pyramids, causal masking, and a pooling decoder, and the paper reports confidence intervals and ablations. However, the entire empirical contribution is conditional on the validity of the intention labels, and the label construction in Section 4.1 is not a valid measure of which finding the radiologist is examining. Because the labels are deterministic functions of report dictation timing and final report content, the reported gains are at risk of merely reproducing the temporal-containment rule used to generate the labels. The paper would be a useful benchmark only after this construct-validity issue is addressed; as it stands, the central claim is not supported.

major comments (4)
  1. [Section 4.1, Eq. (4)] The ground-truth labels in RadExplore are defined by l_ik = 1 if there exists j with tau_i <= s^e_j and k = c_j. This makes the label a deterministic function of the dictation timeline and the report's final content; the fixation's spatial coordinates (x_i, y_i) play no role. A fixation anywhere on the image is labeled as 'intending' every finding that happens to be dictated later, and a fixation before any dictation is labeled with all later-mentioned findings. The same temporal-containment logic underlies RadSeq (Eq. (5)) and RadHybrid (Eq. (6)), and Eq. (6) additionally labels every fixation in the first second as positive for all findings. Consequently, the abstract's claim that the model predicts 'which findings radiologists are examining at specific moments' is not supported by this ground truth. The authors need either to validate these labels against an independent measure of fixation intention (e.g., manual annotation or anatomical localization of fixations relative to the reported findings) or to explicitly reframe the contribution as predicting future dictation mentions rather than cognitive intention.
  2. [Section 4.4, Table 2] No baseline in Table 2 exploits the report/timing prior that is baked into Eqs. (4)-(6). Because the labels are deterministic transforms of the dictation timeline, a trivial predictor that outputs all findings mentioned later in the report (or, for RadHybrid, all ones for fixations with tau_i <= 1) can achieve high recall and substantial F1 without using gaze information. The comparison against MLP, LSTM, Karargyris et al., and ChestSearch therefore cannot establish that the reported ~70% F1 reflects decoding of fixation intention. Please add (a) a report-prior baseline that reconstructs labels directly from dictation sentence end times, (b) an all-positive/global-findings baseline, and (c) a control with shuffled or random fixation coordinates, and report phase-stratified metrics for RadHybrid's initial-scan period. If RadGazeIntent does not clearly beat these controls, the central claim should be withdrawn.
  3. [Section 4.6, Table 3] The ablation results are not evidence for intention decoding. Every removed component (pool attention, temporal/spatial embedding, peripheral features, fovea mapping) also affects the model's ability to predict global report findings from the image. For example, the large drop from removing Fovea Mapping (11.73-12.12 percentage points) is expected if high-resolution visual features are needed to infer any pathology, independent of where the eyes look. The ablations should be repeated on the report-prior and shuffled-location controls, or interpreted only as sensitivity analyses for the proposed architecture, not as support for the claim that the model localizes diagnostic intent.
  4. [Section 5, Limitations] The manuscript's own limitations concede that 'intention is inherently abstract and cannot be directly observed in eye tracking datasets' and that 'not all fixations may map cleanly to a specific diagnostic objective.' This is not a minor caveat: Eqs. (4)-(6) impose a very specific observable proxy, and the model is trained and evaluated against that same proxy. The paper should discuss construct validity explicitly, including why temporal containment relative to dictation is a valid measure of the radiologist's current examination target, and should include an evaluation against an independent measure or an explicit statement that the benchmark measures only this proxy.
minor comments (7)
  1. [Section 4.1, RadSeq] The computation of beg_k and end_k from 'Gazesearch's constraints' is not specified; the reader cannot reconstruct RadSeq labels from the text. Please give the exact algorithm or a reference with the full equation.
  2. [Sections 4.2 and 4.3] Training details are inconsistent: Section 4.2 states Adam with learning rate 1e-4 for 100 epochs, while Section 4.3 states AdamW with learning rate 1e-5 for 4,000 iterations. Clarify which configuration produced Table 2.
  3. [Section 4.4] The decision threshold for accuracy, F1, precision, and recall is not stated; Figure 5 uses a confidence threshold of 0.5, but this should be stated explicitly for the quantitative metrics or replaced with threshold-free metrics such as average precision.
  4. [Section 4.1] The number of intention classes K and the CheXbert mapping procedure are not fully described; Figure 4 lists 13 findings, but the exact class set and the handling of sentences with no finding or multiple findings should be specified.
  5. [Throughout] Please fix typos such as 'designe', 'wehre', 'Aritificial', 'objecive', and the double semicolon in the Limitations paragraph.
  6. [Figure 5] Figure 5's caption says green and red points, but the figure uses check and cross markers; align the caption with the actual visualization.
  7. [Introduction and Datasets] As a benchmark contribution, the paper should state whether RadSeq, RadExplore, and RadHybrid will be released and provide code links; otherwise, the new datasets are not reproducible from the description alone.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the label-construction concern is a construct-validity issue, not a reduction of the prediction to its inputs.

full rationale

The paper's central claim is an empirical benchmark result: RadGazeIntent is trained to predict intention labels defined by Eqs. 4–6, which are derived from dictation timing and report findings. The model's inputs are the chest X-ray image and fixation features; the label rule (e.g., 'fixation occurs before a sentence that mentions finding k') is not provided to the model as an input. Thus the prediction is not equivalent to its inputs by construction. A model that ignores gaze position and simply predicts findings present in the image could potentially score well on these labels, and the paper does not evaluate such a report-prior or temporal-prior baseline; this is a legitimate experimental and construct-validity concern. The paper's own limitation section acknowledges that 'intention is inherently abstract and cannot be directly observed in eye tracking datasets' and that 'not all fixations may map cleanly to a specific diagnostic objective.' However, this is an issue of whether the labels measure true cognitive intent, not a circular derivation: the model could fail to predict the labels, and the reported improvements over the gaze-based baselines are empirical. The only self-citation, GazeSearch [39], is used to define RadSeq's temporal constraints, but it is not invoked as a proof, uniqueness theorem, or forced ansatz, and it does not determine the model's performance. The derivation chain is therefore self-contained relative to the paper's stated task, with the caveat that the task's construct validity is external to the paper's equations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that the temporal patterns of report dictation can stand in for the radiologist's current intention. This is a domain assumption with no independent validation, and it drives the label construction (Eq. 4).

free parameters (3)
  • initial_scan_time_tau_star = 1 second
    Hand-chosen constant defining the broad-overview window in RadHybrid labels (Eq. 6). Results would change with a different value.
  • gazesearch_radius = 0
    Radius set to zero when adapting GazeSearch constraints to build RadSeq, discarding the original filtering and changing which fixations fall in each intention interval.
  • confidence_threshold = 0.5
    Threshold used in qualitative evaluation (Figure 5) to mark a prediction correct; not used in quantitative metrics.
assumptions (4)
  • domain assumption Eye movements provide a window into cognitive processes and diagnostic intention.
    Stated in the Introduction and Discussion; the entire task relies on this premise.
  • domain assumption Radiologist intention is captured by three patterns: systematic sequential search, uncertainty-driven exploration, and hybrid behavior.
    Section 1; the three datasets are built around these three constructs.
  • domain assumption CheXbert correctly extracts the set of findings mentioned in each dictated report sentence.
    Section 4.1; labels are derived from CheXbert output.
  • ad hoc to paper A fixation's intention label can be inferred from the time of report dictation, specifically that a fixation before a finding is dictated is 'intending' that finding.
    Eq. (4) in section 4.1; this is the paper's own operationalization and is not validated against any independent measure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpreting Radiologist's Intention from Eye Movements in Chest X-ray Diagnosis." pith.science (2026). https://pith.science/paper/WIQINTAL

@misc{pith2026250712461,
  author       = {Pith},
  title        = {Pith review of: Interpreting Radiologist's Intention from Eye Movements in Chest X-ray Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WIQINTAL}},
  note         = {Machine review of arXiv:2507.12461}
}
read the original abstract

Radiologists rely on eye movements to navigate and interpret medical images. A trained radiologist possesses knowledge about the potential diseases that may be present in the images and, when searching, follows a mental checklist to locate them using their gaze. This is a key observation, yet existing models fail to capture the underlying intent behind each fixation. In this paper, we introduce a deep learning-based approach, RadGazeIntent, designed to model this behavior: having an intention to find something and actively searching for it. Our transformer-based architecture processes both the temporal and spatial dimensions of gaze data, transforming fine-grained fixation features into coarse, meaningful representations of diagnostic intent to interpret radiologists' goals. To capture the nuances of radiologists' varied intention-driven behaviors, we process existing medical eye-tracking datasets to create three intention-labeled subsets: RadSeq (Systematic Sequential Search), RadExplore (Uncertainty-driven Exploration), and RadHybrid (Hybrid Pattern). Experimental results demonstrate RadGazeIntent's ability to predict which findings radiologists are examining at specific moments, outperforming baseline methods across all intention-labeled datasets.

Figures

Figures reproduced from arXiv: 2507.12461 by the authors.

Figure 1
Figure 1. From existing gaze datasets (e.g., EGD and REFLACX) that are illustrated in the left image, existing research on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. First, intention may manifest as a systematic sequential [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall framework of RadGazeIntent. RadGazeIntent aims to analyze medical image fixation patterns to determine [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Illustration of our dataset creation process, which [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of radiologist intention prediction results across three datasets and various intention classes. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 45 canonical work pages

  1. [1]

    Bahar Aydemir, Ludo Hoffstetter, Tong Zhang, Mathieu Salzmann, and Sabine Susstrunk. 2023. TempSAL - Uncovering Temporal Information for Deep Saliency Prediction.. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2023

  2. [2]

    Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, et al. 2021. Big self-supervised models advance medical image classifica- tion. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 3478–3488

    work page 2021

  3. [3]

    Raymond Bertram, Laura Helle, Johanna K Kaakinen, and Erkki Svedström. 2013. The effect of expertise on eye movement behaviour in medical image perception. PloS one 8, 6 (2013), e66169

    work page 2013

  4. [4]

    Moinak Bhattacharya, Shubham Jain, and Prateek Prasanna. 2022. Gazeradar: A gaze and radiomics-guided disease localization framework. In International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 686–696

    work page 2022

  5. [5]

    Ricardo Bigolin Lanfredi, Mingyuan Zhang, William F Auffermann, Jessica Chan, Phuong-Anh T Duong, Vivek Srikumar, Trafton Drew, Joyce D Schroeder, and Tolga Tasdizen. 2022. REFLACX, a dataset of reports and eye-tracking data for localization of abnormalities in chest x-rays. Scientific data 9, 1 (2022), 350

    work page 2022

  6. [6]

    Souradeep Chakraborty and otherss. 2022. Predicting Visual Attention in Graphic Design Documents. IEEE Transactions on Multimedia (TMM) (2022)

    work page 2022

  7. [7]

    Shi Chen, Ming Jiang, Jinhui Yang, and Qi Zhao. 2020. AiR: Attention with Reasoning Capability.. In Proceedings of the European Conference on Computer Vision (ECCV)

    work page 2020

  8. [8]

    Shi Chen, Nachiappan Valliappan, Shaolei Shen, Xinyu Ye, Kai Kohlhoff, and Junfeng He. 2023. Learning from Unique Perspectives: User-aware Saliency Modeling.. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2023

Show all 60 references
  1. [9]

    Xianyu Chen, Ming Jiang, and Qi Zhao. 2021. Predicting Human Scanpaths in Visual Question Answering.. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2021

  2. [10]

    Xianyu Chen, Ming Jiang, and Qi Zhao. 2024. Beyond Average: Individualized Visual Scanpath Prediction.. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2024

  3. [11]

    Xianyu Chen, Ming Jiang, and Qi Zhao. 2024. Beyond Average: Individualized Vi- sual Scanpath Prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 25420–25431

    2024

  4. [12]

    Yupei Chen et al . 2022. Characterizing Target-Absent Human Attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 5031–5040

    2022

  5. [13]

    Zhenzhong Chen and Wanjie Sun. 2018. Scanpath Prediction for Visual Attention using IOR-ROI LSTM.. In International Joint Conference on Artificial Intelligence (IJCAI)

    2018

  6. [14]

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. 2022. Masked-attention mask transformer for universal image segmen- tation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1290–1299

    2022

  7. [15]

    Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. 2018. Predicting Human Eye Fixations via an LSTM-based Saliency Attentive Model. IEEE Transactions on Image Processing (IEEE TIP) (2018)

    2018

  8. [16]

    Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. 2021. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision . 6824– 6835

    2021

  9. [17]

    Camilo Fosco, Vincent Casser, Amish Kumar Bedi, Peter O’Donovan, Aaron Hertz- mann, and Zoya Bylinskii. 2020. Predicting Visual Importance Across Graphic Design Types.. In ACM Symposium on User Interface Software and Technology

    2020

  10. [18]

    Matej Gazda, Ján Plavka, Jakub Gazda, and Peter Drotar. 2021. Self-supervised deep convolutional neural network for chest x-ray classification. IEEE Access 9 (2021), 151972–151982

    2021

  11. [19]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep residual learning for image recognition. arXiv e-prints. arXiv preprint arXiv:1512.03385 10 (2015)

    2015 arXiv

  12. [20]

    Chihcheng Hsieh, André Luís, José Neves, Isabel Blanco Nobre, Sandra Costa Sousa, Chun Ouyang, Joaquim Jorge, and Catarina Moreira. 2024. EyeXNet: Enhancing Abnormality Detection and Diagnosis via Eye-Tracking and X-ray Fusion. Machine Learning and Knowledge Extraction 6, 2 (2...

    2024

  13. [21]

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger

  14. [22]

    Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. 2015. SALICON: Reducing the Semantic Gap in Saliency Prediction by Adapting Deep Neural Networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV)

    2015

  15. [23]

    Jeremy Irvin et al. 2019. Chexpert: A large chest radiograph dataset with uncer- tainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, Vol. 33. 590–597

    2019

  16. [24]

    Sen Jia and Neil D. B. Bruce. 2020. EML-NET:An Expandable Multi-Layer NET- work for Saliency Prediction. Image and Vision Computing (2020)

    2020

  17. [25]

    Alexandros Karargyris et al. 2021. Creation and validation of a chest X-ray dataset with eye-tracking and report dictation for AI development. Scientific Data 8, 1 (2021), 1–18

    2021

  18. [26]

    Matthias Kümmerer, Matthias Bethge, and Thomas S. A. Wallis. 2022. DeepGaze III: Modeling free-viewing human scanpaths with deep learning.Journal of Vision (JoV) (2022)

    2022

  19. [27]

    Matthias Kümmerer, Thomas S. A. Wallis, and Matthias Bethge. 2016. DeepGaze II: Reading fixations from deep features trained on object recognition. arXiv preprint arXiv:1610.01563 (2016)

    2016 arXiv

  20. [28]

    Peizhao Li, Junfeng He, Gang Li, Rachit Bhargava, Shaolei Shen, Nachiappan Valli- appan, Youwei Liang, Hongxiang Gu, Venky Ramachandran, Golnaz Farhadi, Yang Li, Kai J Kohlhoff, and Vidhya Navalpakkam. 2023. UniAR: Unifying Human Atten- tion and Response Prediction on Visual C...

    2023 arXiv

  21. [29]

    Zhe Li, Chong Wang, Mei Han, Yuan Xue, Wei Wei, Li-Jia Li, and Li Fei-Fei

  22. [30]

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. 2017. Feature pyramid networks for object detection. In Proceed- ings of the IEEE conference on computer vision and pattern recognition . 2117–2125

    2017

  23. [31]

    Fengbei Liu, Yu Tian, Filipe R Cordeiro, Vasileios Belagiannis, Ian Reid, and Gustavo Carneiro. 2021. Self-supervised mean teacher for semi-supervised chest x-ray classification. In Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunctio...

    2021

  24. [32]

    Jingyu Liu, Gangming Zhao, Yu Fei, Ming Zhang, Yizhou Wang, and Yizhou Yu. 2019. Align, attend and locate: Chest x-ray diagnosis via contrast induced attention network with limited supervision. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision . 10632–10641

    2019

  25. [33]

    Quande Liu, Lequan Yu, Luyang Luo, Qi Dou, and Pheng Ann Heng. 2020. Semi- supervised medical image classification with relation-driven self-ensembling model. IEEE transactions on medical imaging 39, 11 (2020), 3429–3440

    2020

  26. [34]

    I Loshchilov. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    2017 arXiv

  27. [35]

    Sounak Mondal et al. 2023. Gazeformer: Scalable, Effective and Fast Prediction of Goal-Directed Human Attention.. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2023

  28. [36]

    José Neves, Chihcheng Hsieh, Isabel Blanco Nobre, Sandra Costa Sousa, Chun Ouyang, Anderson Maciel, Andrew Duchowski, Joaquim Jorge, and Catarina Moreira. 2024. Shedding light on ai in radiology: A systematic review and taxonomy of eye gaze-driven interpretability in deep lear...

    2024

  29. [37]

    Peixi Peng, Wanshu Fan, Yue Shen, Wenfei Liu, Xin Yang, Qiang Zhang, Xiaopeng Wei, and Dongsheng Zhou. 2024. Eye gaze guided cross-modal alignment network for radiology report generation.IEEE Journal of Biomedical and Health Informatics (2024)

    2024

  30. [38]

    Trong Thang Pham, Jacob Brecheisen, Anh Nguyen, Hien Nguyen, and Ngan Le

  31. [39]

    Trong Thang Pham, Tien-Phat Nguyen, Yuki Ikebe, Akash Awasthi, Zhigang Deng, Carol C Wu, Hien Nguyen, and Ngan Le. 2024. GazeSearch: Radiology Findings Search Benchmark. arXiv preprint arXiv:2411.05780 (2024)

    2024 arXiv

  32. [40]

    Mengyu Qiu, Yi Guo, Mingguang Zhang, Jingwei Zhang, Tian Lan, and Zhilin Liu. 2023. Simulating Human Visual System Based on Vision Transformer.. In Proceedings of the 2023 ACM Symposium on Spatial User Interaction

    2023

  33. [41]

    Pranav Rajpurkar, Jeremy Irvin, Kaylie Zhu, Brandon Yang, Hershel Mehta, Tony Duan, Daisy Ding, Aarti Bagul, Curtis Langlotz, Katie Shpanskaya, et al

  34. [42]

    Ng, and Matthew P

    Akshay Smit, Saahil Jain, Pranav Rajpurkar, Anuj Pareek, Andrew Y. Ng, and Matthew P. Lungren. 2020. CheXbert: Combining Automatic Labelers and Expert Annotations for Accurate Radiology Report Labeling Using BERT. arXiv:2004.09167 [cs.CL]

    2020 arXiv

  35. [43]

    Wanjie Sun, Zhenzhong Chen, and Feng Wu. 2019. Visual Scanpath Prediction using IOR-ROI Recurrent Mixture Density Network. IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI) (2019)

    2019

  36. [44]

    Sina Taslimi, Soroush Taslimi, Nima Fathi, Mohammadreza Salehi, and Moham- mad Hossein Rohban. 2022. SwinCheX: Multi-label classification on chest X-ray images with transformers. arXiv preprint arXiv:2206.04246 (2022)

    2022 arXiv

  37. [45]

    arXiv preprint arXiv:1711.05225 (2017)

    Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225 (2017)

    2017 arXiv

  38. [46]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017)

    2017

  39. [47]

    Fuying Wang, Yuyin Zhou, Shujun Wang, Varut Vardhanabhuti, and Lequan Yu

  40. [48]

    Joy T Wu, Nkechinyere N Agu, Ismini Lourentzou, Arjun Sharma, Joseph A Paguio, Jasper S Yao, Edward C Dee, William Mitchell, Satyananda Kashyap, Andrea Giovannini, et al. 2021. Chest imagenome dataset for clinical reasoning. arXiv preprint arXiv:2108.00316 (2021)

    2021 arXiv

  41. [49]

    Tom van Sonsbeek, Xiantong Zhen, Dwarikanath Mahapatra, and Marcel Worring

  42. [50]

    Chaochao Yan, Jiawen Yao, Ruoyu Li, Zheng Xu, and Junzhou Huang. 2018. Weakly supervised deep learning for thoracic disease classification and localiza- tion on chest x-rays. In Proceedings of the 2018 ACM international conference on bioinformatics, computational biology, and ...

    2018

  43. [51]

    Zhibo Yang et al. 2020. Predicting Goal-directed Human Attention Using Inverse Reinforcement Learning.. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2020

  44. [52]

    Zhibo Yang, Sounak Mondal, Seoyoung Ahn, Gregory Zelinsky, Minh Hoai, and Dimitris Samaras. 2022. Target-absent Human Attention.. In Proceedings of the European Conference on Computer Vision (ECCV)

    2022

  45. [53]

    Zhibo Yang, Sounak Mondal, Seoyoung Ahn, Gregory Zelinsky, Minh Hoai, and Dimitris Samaras. 2023. Predicting Human Attention using Computational Attention. arXiv preprint arXiv:2303.09383v2 (2023)

    2023 arXiv

  46. [54]

    Li Yao, Jordan Prosky, Eric Poblenz, Ben Covington, and Kevin Lyman. 2018. Weakly supervised medical diagnosis and localization from multiple resolutions. arXiv preprint arXiv:1803.07703 (2018)

    2018 arXiv

  47. [55]

    Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. 2020. Layoutlm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining . 1192–1200

    2020

  48. [2017]

    In Proceedings of the IEEE conference on computer vision and pattern recognition

    Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition . 4700–4708

  49. [2018]

    In Proceedings of the IEEE conference on computer vision and pattern recognition

    Thoracic disease identification and localization with limited supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition . 8290–8299

  50. [2022]

    Advances in Neural Information Processing Systems 35 (2022), 33536–33549

    Multi-granularity cross-modal alignment for generalized medical visual representation learning. Advances in Neural Information Processing Systems 35 (2022), 33536–33549

    2022

  51. [2023]

    In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Probabilistic Integration of Object Level Annotations in Chest X-ray Classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 3630–3640. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trong Thang Pham, Anh Nguyen, Zhigang De...

    2018

  52. [2024]

    InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    I-AI: A Controllable & Interpretable AI System for Decoding Radiologists’ Intense Focus for Accurate CXR Diagnoses. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 7850–7859

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.