Pith. sign in

REVIEW 5 major objections 7 minor 1 cited by

GazeLT: Visual attention-guided long-tailed disease classification in chest radiographs

T0 review · 5 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Splitting a radiologist's gaze into four successive time windows preserves rare-disease cues that static attention maps lose, and distilling those windowed maps into a classifier raises tail-class accuracy by roughly 11% on two chest-X-ray

desk verdict Temporal gaze windowing is a real novelty, but the paper overclaims: the MIMIC imbalanced result contradicts the abstract, and the gaze contribution is not isolated from plain distillation. read the letter →

arxiv 2508.09478 v2 pith:BES5IDSP submitted 2025-08-13 cs.CV

Moinak Bhattacharya , Gagandeep Singh , Shubham Jain , Prateek Prasanna This is my paper · ORCID
classification cs.CV
keywords eyegazelong-tailedclassificationchestradiographtemporalvisualattentionintegrationanddisintegrationknowledgedistillationteacher-studentframeworkclassimbalance
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

Chest X-ray datasets are sharply imbalanced: common diseases occupy the head of the distribution while rare ones, such as pneumoperitoneum, hernia, and pneumomediastinum, can appear in fewer than a hundred training images, and standard losses leave those tail classes largely unlearnable. GazeLT's bet is that a radiologist's eye movements contain the missing signal: readers first fixate on obvious abnormalities and later scan for incidental findings, so the temporal order of gaze, not the duration-averaged gaze map, is where rare-disease information lives. The method cuts each gaze record into four equal time windows, builds a fine-grained 'integration' attention map and a coarse 'disintegration' attention map for each window, trains transformer teachers to reproduce those maps, and distills the fused features into a ResNet-50 student trained with the LDAM long-tailed loss. Sympathetically read, the paper establishes that this temporal-attention teaching signal raises average accuracy on the balanced test sets to 0.410 on NIH-CXR-LT and 0.418 on MIMIC-CXR-LT, with tail-class gains of 10.9% and 12.2% over the strongest baselines at a roughly 5% cost in head-class accuracy.

What carries the argument

The load-bearing object is the Time-Windowed Integration–Disintegration pair (TW-I/TW-D) embedded in a teacher–student loop. Radiologist fixation points are partitioned into $n=4$ duration-equal windows; within each window, integration attention selects fixations inside a distance threshold (set by K-means clustering) to model focused inspection of a single region, while disintegration attention smooths all raw fixations with a wide Gaussian to model coarse global scanning. Two transformer teachers, a FocalNet for integration and a GC-ViT for disintegration, are trained with a normalized $\ell^2$ Visual Attention Loss to emit these per-window maps, and the fused teacher features are transfer

What would settle it

Train the identical teacher–student pipeline with (a) the static aggregated gaze map replicated across all four time windows, and (b) per-window maps whose window labels are randomly permuted across training images. If balanced tail-class accuracy does not drop materially under either variant, the temporal-decomposition claim is not what drives the gain. A direct check of the mechanism: measure whether fixations within each window preferentially fall inside ground-truth bounding boxes of specific classes; the claimed head-then-tail reading order predicts that windows should separate head- and

Watch

Extended reading notes

Core claim

The central claim is that the temporal structure of a radiologist's visual search is class-discriminative in a way that a static, duration-averaged attention map is not, and that this structure is especially informative for the rare, long-tailed disease classes. To make the claim operational, the paper partitions gaze fixations into $n=4$ equal time windows and, within each window, computes two kinds of attention from the raw fixation points: integration attention, which thresholds and clusters fixations to isolate the region of focused inspection (the 'main' region plus substitutes), and disintegration attention, a broad Gaussian over all raw fixations that models global scanning. Two trans

Load-bearing premise

The load-bearing premise is that four-window gaze maps, learned from only 1,083 to 2,616 gaze-annotated images, separate head-class from tail-class information that the static averaged gaze map does not; if the per-window maps are noisy or temporally redundant, the reported gains would come from the distillation setup rather than from gaze timing.

Editorial extensions

If this is right

  • On the balanced test sets, GazeLT reaches 0.410 average accuracy on NIH-CXR-LT and 0.418 on MIMIC-CXR-LT, versus 0.369 and 0.387 for the best long-tailed-loss baseline (Decoupling-cRT).
  • Per-class accuracy on tail diseases rises 10.9% (NIH-CXR-LT) and 12.2% (MIMIC-CXR-LT) over the baselines, with roughly a 5% drop in head-class accuracy, a trade the paper argues is clinically preferable because tail findings are often the serious ones.
  • The temporal windowing itself is the active ingredient: TW-I+TW-D outperforms the static I+D variant by 4.7% in balanced accuracy, and the $n=4$ configuration beats $n=2$ (0.382) and $n=8$ (0.403) on the NIH-CXR-LT balanced set.
  • Gaze recordings are needed only to train the teacher; at inference the student classifies from the chest X-ray alone, so deployment requires no eye-tracking hardware.
  • In multi-class evaluation on NIH-CXR-LT, GazeLT reaches balanced accuracy 0.421 and AUC 0.829, clearly above the static-gaze baselines (0.187–0.193 balanced accuracy, roughly 0.79 AUC).

Reading between the lines

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

  • An untested alternative explanation: because the teacher is trained on only 1,083–2,616 gaze-annotated images (from the EGD-CXR and REFLACX collections), part of the student gain could come from the distillation regularizer itself rather than from the temporal structure of gaze; a window-permutation or static-replication ablation would isolate the mechanism.
  • The $n=4$ optimum suggests the window count should track the duration of a reading episode; if gaze is collected under shorter or longer viewing protocols, the optimal $n$ would be expected to move, which is a testable consequence for other gaze datasets.
  • The same head-then-tail scan hypothesis should transfer to other long-tailed medical tasks, including pathology, retinal imaging, and 3D modalities, where experts also scan for incidental findings; the paper only demonstrates chest X-rays.
  • Because head-class accuracy drops about 5% while tail-class accuracy rises about 11%, GazeLT shifts the classifier operating point; whether that trade is net-positive in practice depends on the clinical cost of missing a rare finding versus over-calling a common one, which the paper does not quantify.
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

5 major / 7 minor

Summary. GazeLT proposes a teacher–student framework for long-tailed thoracic disease classification that incorporates radiologists' eye-gaze patterns. A two-branch transformer teacher (TW-I and TW-D) is trained on time-windowed integration/disintegration heatmaps derived from gaze fixations, using custom temporal visual attention losses. The frozen teacher distills fused features to a ResNet-50 student via a Bhattacharyya-distance loss, while the student is also trained with LDAM. The authors evaluate on NIH-CXR-LT and MIMIC-CXR-LT, reporting balanced-set average accuracy and imbalanced-set balanced accuracy, with ablations over static vs temporal attention and over the number of time windows. The central claim is that temporal decomposition of gaze into integration and disintegration windows improves long-tailed classification beyond both long-tailed losses and prior gaze-guided baselines.

Significance. If the causal contribution of temporal gaze were established, this would be a novel and clinically plausible prior for long-tailed medical image classification, with the practical advantage that gaze is needed only at training time. The paper releases code and uses public benchmarks, which aids reproducibility. The tail-class results are potentially important. However, the current evaluation does not isolate the gaze-temporal mechanism from general feature distillation or from the added LDAM loss, and one headline quantitative claim is contradicted by the reported MIMIC imbalanced-test number. The idea is worth pursuing, but the evidence presented does not yet support the central attribution.

major comments (5)
  1. [Abstract; §5.2, Table 1] The abstract states GazeLT outperforms the best long-tailed loss by 4.1% in average accuracy metrics. On the MIMIC-CXR-LT imbalanced test set, GazeLT has bAcc 0.292 while Decoupling-cRT has 0.296 (Table 1); GazeLT is therefore below the best long-tailed baseline on a reported primary metric. The 4.1% improvement holds only for the NIH-CXR-LT balanced set (0.410 vs 0.369), and the corresponding MIMIC balanced-set improvement is 0.031. Please restrict the claim to the specific settings/datasets, or report dataset-wise results in the abstract.
  2. [§3.3, §5.3, Tables 2–3] The paper's central causal claim—that temporal gaze integration/disintegration drives the gains—is not isolated. The student objective L_s = L_LDAM + λL_BD (Eqs. 3–4) includes feature distillation from a frozen transformer teacher; distillation from a stronger transformer teacher to a CNN is a known regularizer independent of the semantic content distilled. The static I, D, and I+D ablations in Table 2 still use gaze and distillation, so they cannot separate 'gaze content' from 'distillation itself.' A no-gaze teacher control (same architecture, trained on the same CXR images without gaze supervision, or with randomized attention maps) is required before the temporal mechanism can be credited with the observed 0.363→0.410 gain. The MIMIC ablation in Table 3 does not even include an I+D row, so temporal contribution is untested there.
  3. [§5.2, Table 1; §5.4, Table 4] The comparison against the visual-attention baselines is confounded by the long-tailed loss. GazeLT is trained with L_LDAM + λL_BD, while RadioTransformer and GazeRadar are reported with their original (standard) objectives; their very low average accuracies (0.19 on NIH balanced set) are consistent with no LT rebalancing. Thus the headline 21.7% gap (and Table 4) conflates gaze guidance with the effect of LDAM. Please add versions of RadioTransformer and GazeRadar trained with LDAM (or otherwise exactly matched), and/or report GazeLT without LDAM, so the attention contribution is isolated.
  4. [§3.1 vs §4.2] The heatmap-generation parameters are stated inconsistently: §3.1 says integration uses σ=64 and disintegration uses σ=128, while §4.2 says σ=128 for integration and σ=64 for disintegration, after a grid search. Since these heatmaps define the teacher supervision targets (Eqs. 1–2), the discrepancy is not cosmetic. Please state which values were actually used and ensure all sections and the appendix agree.
  5. [§5.5, Table 5] The number of time windows n=4 is selected by comparing average accuracy on the NIH-CXR-LT balanced test set (Table 5), not on a held-out validation split; reporting a hyperparameter tuned on the test set inflates the result. No analogous sensitivity analysis is given for MIMIC-CXR-LT, so the claim that n=4 is optimal across datasets is unsupported. Please use a validation split for model selection and report all n for both datasets with confidence intervals.
minor comments (7)
  1. [§5.2] There is a duplicated phrase: 'It is important to note that the proposed method It is important to note that...'.
  2. [Eqs. (1)–(2)] The loss sums over i=1..L but the quantities are indexed by t; define L and t clearly (number of layers/windows?) and make the index consistent.
  3. [Abstract / Appendix Figure 6] MIMIC-CXR-LT sample size appears as n=111898 in the abstract and n=111792 in Figure 6; unify the counts.
  4. [Data availability] 'EDG-CXR' is presumably 'EGD-CXR' used in §4.1; fix the typo.
  5. [§5.5, Figures 7–8] References to 'Rebuttal Figure 7/8' should be 'Figure 7/8'.
  6. [§4.5, Tables 1–3] T-tests and p-values are mentioned, but no standard deviations, confidence intervals, or number of seeds are reported; add these so the 'significant' claims can be assessed.
  7. [§3.3] The distillation loss Eq. (3) is written with f_s and J(f_I,f_D) before explaining normalization; please clarify the exact feature-normalization and projection steps and the value of λ.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the gaze-supervised teacher provides external supervision and the reported gains are empirical.

full rationale

GazeLT's derivation chain is an empirical teacher-student pipeline, not a closed-form derivation. The teacher's temporal attention targets f_I^t and f_D^t are constructed from raw eye-gaze fixations (EGD-CXR/REFLACX) via Algorithm 1 and time-window binning (Section 3.1), i.e., from externally collected gaze ground truth, not from the disease labels or the student's predictions. The teacher is trained with L_I-tVAL and L_D-tVAL (Eqs. 1-2) to regress those gaze-derived heatmaps, and the student is trained with L_s = L_LDAM + λ L_BD (Eqs. 3-4), where L_BD aligns student features to the frozen teacher. The central claim—that temporally windowed gaze improves long-tailed classification—is tested on held-out NIH-CXR-LT and MIMIC-CXR-LT labels (Tables 1-3), including comparisons to the authors' own static-gaze methods as baselines. No equation defines the teacher's features in terms of the student's labels, and no fitted parameter is renamed as a prediction: the choices of n=4 and the Gaussian sigmas are model-selection decisions (Sections 4.2 and 5.5) rather than a definitional equivalence. The self-citations to GazeRadar/RadioTransformer serve as empirical baselines, not as the justification for the core claim. The absence of a no-gaze teacher control is a validity concern, but it does not make the derivation circular because the supervision signal originates from external eye-tracking data.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The method relies on untested domain assumptions about gaze transfer and per-window class information, and on several hand-picked hyperparameters, but introduces no invented entities. Free parameters (n, sigmas, k, lambda) are tuned on the reported datasets, so headline improvements are partly selected.

free parameters (5)
  • number of time windows n = 4
    Selected via sensitivity analysis over n in {2,4,8} on NIH-CXR-LT balanced test (Table 5); all main results use n=4.
  • Gaussian sigma for integration attention = 128
    Selected through validation grid search over {32,64,96,128,160} for integration heatmap smoothing (Section 4.2).
  • Gaussian sigma for disintegration attention = 64
    Selected through validation grid search over {32,64,96,128,160} for disintegration heatmap smoothing (Section 4.2).
  • cluster count k in K-means = 2
    Fixation points clustered into k=2; the lower-variance cluster is treated as focal attention (Algorithm 1, Section 3.1).
  • distillation weight lambda = not reported
    Student loss Ls = L_LDAM + lambda * L_BD (Section 3.3); the value of lambda is not stated in the paper.
assumptions (3)
  • domain assumption Per-window gaze fixation maps are learnable from images and carry distinct class-relevant information for head and tail classes.
    Central hypothesis motivating teacher pretraining; enters in Sections 3.1 and 3.2 and Figure 4; not validated by any no-gaze control.
  • domain assumption Gaze data collected on MIMIC-CXR images (EGD-CXR, REFLACX) transfers to NIH-CXR-LT images.
    Teacher is trained on MIMIC gaze images and frozen; student is evaluated on NIH-CXR-LT (Section 4.1).
  • standard math Standard math operations (Gaussian smoothing, K-means, Bhattacharyya distance) used without formal justification.
    Used in Algorithm 1 and Eq. 3; conventional and not load-bearing beyond choice of parameters.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GazeLT: Visual attention-guided long-tailed disease classification in chest radiographs." pith.science (2026). https://pith.science/paper/BES5IDSP

@misc{pith2026250809478,
  author       = {Pith},
  title        = {Pith review of: GazeLT: Visual attention-guided long-tailed disease classification in chest radiographs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BES5IDSP}},
  note         = {Machine review of arXiv:2508.09478}
}
read the original abstract

In this work, we present GazeLT, a human visual attention integration-disintegration approach for long-tailed disease classification. A radiologist's eye gaze has distinct patterns that capture both fine-grained and coarser level disease related information. While interpreting an image, a radiologist's attention varies throughout the duration; it is critical to incorporate this into a deep learning framework to improve automated image interpretation. Another important aspect of visual attention is that apart from looking at major/obvious disease patterns, experts also look at minor/incidental findings (few of these constituting long-tailed classes) during the course of image interpretation. GazeLT harnesses the temporal aspect of the visual search process, via an integration and disintegration mechanism, to improve long-tailed disease classification. We show the efficacy of GazeLT on two publicly available datasets for long-tailed disease classification, namely the NIH-CXR-LT (n=89237) and the MIMIC-CXR-LT (n=111898) datasets. GazeLT outperforms the best long-tailed loss by 4.1% and the visual attention-based baseline by 21.7% in average accuracy metrics for these datasets. Our code is available at https://github.com/lordmoinak1/gazelt.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Pathologist Attention-Aligned Report Generation for Prostate Histopathology

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Using pathologists' eye movements as a training signal improves prostate pathology report generation and makes model attention more human-like.

Reference graph

Works this paper leans on

54 extracted references · 51 canonical work pages · cited by 1 Pith paper

  1. [1]

    Activity segmentation and identification based on eye gaze features

    Sabrina Amrouche, Benedikt Gollan, Alois Ferscha, and Josef Heftberger. Activity segmentation and identification based on eye gaze features. In Proceedings of the 11th Pervasive Technologies Related to Assistive Environments Conference, pages 75--82, 2018

    work page 2018

  2. [2]

    Gazediff: A radiologist visual attention guided diffusion model for zero-shot disease classification

    Moinak Bhattacharya and Prateek Prasanna. Gazediff: A radiologist visual attention guided diffusion model for zero-shot disease classification. In Medical Imaging with Deep Learning, 2024

    work page 2024

  3. [3]

    Gazeradar: A gaze and radiomics-guided disease localization framework

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

    work page 2022

  4. [4]

    Radiotransformer: a cascaded global-focal transformer for visual attention--guided disease classification

    Moinak Bhattacharya, Shubham Jain, and Prateek Prasanna. Radiotransformer: a cascaded global-focal transformer for visual attention--guided disease classification. In European Conference on Computer Vision, pages 679--698. Springer, 2022 b

    work page 2022

  5. [5]

    RadGazeGen: Radiomics and Gaze-guided Medical Image Generation using Diffusion Models

    Moinak Bhattacharya, Gagandeep Singh, Shubham Jain, and Prateek Prasanna. Radgazegen: Radiomics and gaze-guided medical image generation using diffusion models. arXiv preprint arXiv:2410.00307, 2024

    work page Pith review arXiv 2024

  6. [6]

    Learning imbalanced datasets with label-distribution-aware margin loss

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019

    2019

  7. [7]

    Human gaze-aware attentive object detection for ambient intelligence

    Dae-Yong Cho and Min-Koo Kang. Human gaze-aware attentive object detection for ambient intelligence. Engineering Applications of Artificial Intelligence, 106: 0 104471, 2021

    work page 2021

  8. [8]

    The potential and value of objective eye tracking in the ophthalmology clinic

    Rosie Clark, James Blundell, Matt J Dunn, Jonathan T Erichsen, Mario E Giardini, Irene Gottlob, Chris Harris, Helena Lee, Lee Mcilreavy, Andrew Olson, et al. The potential and value of objective eye tracking in the ophthalmology clinic. Eye, 33 0 (8): 0 1200--1202, 2019

    work page 2019

Show all 54 references
  1. [9]

    Activity recognition using eye-gaze movements and traditional interactions

    Fran c ois Courtemanche, Esma A \" meur, Aude Dufresne, Mehdi Najjar, and Franck Mpondo. Activity recognition using eye-gaze movements and traditional interactions. Interacting with Computers, 23 0 (3): 0 202--213, 2011

    2011

  2. [10]

    Class-balanced loss based on effective number of samples

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268--9277, 2019

    2019

  3. [11]

    Understanding attentional modulation of binocular rivalry: a framework based on biased competition

    Kevin Conrad Dieter and Duje Tadin. Understanding attentional modulation of binocular rivalry: a framework based on biased competition. Frontiers in Human Neuroscience, 5: 0 155, 2011

    2011

  4. [12]

    One visual search, many memory searches: An eye-tracking investigation of hybrid search

    Trafton Drew, Sage EP Boettcher, and Jeremy M Wolfe. One visual search, many memory searches: An eye-tracking investigation of hybrid search. Journal of vision, 17 0 (11): 0 5--5, 2017

    2017

  5. [13]

    The foundations of cost-sensitive learning

    Charles Elkan. The foundations of cost-sensitive learning. In International joint conference on artificial intelligence, volume 17, pages 973--978. Lawrence Erlbaum Associates Ltd, 2001

    2001

  6. [14]

    A multiple resampling method for learning from imbalanced data sets

    Andrew Estabrooks, Taeho Jo, and Nathalie Japkowicz. A multiple resampling method for learning from imbalanced data sets. Computational intelligence, 20 0 (1): 0 18--36, 2004

    2004

  7. [15]

    Clinical artificial intelligence quality improvement: towards continual monitoring and updating of ai algorithms in healthcare

    Jean Feng, Rachael V Phillips, Ivana Malenica, Andrew Bishara, Alan E Hubbard, Leo A Celi, and Romain Pirracchio. Clinical artificial intelligence quality improvement: towards continual monitoring and updating of ai algorithms in healthcare. NPJ digital medicine, 5 0 (1): 0 66, 2022

    2022

  8. [16]

    Interactive few-shot learning: Limited supervision, better medical image segmentation

    Ruiwei Feng, Xiangshang Zheng, Tianxiang Gao, Jintai Chen, Wenzhe Wang, Danny Z Chen, and Jian Wu. Interactive few-shot learning: Limited supervision, better medical image segmentation. IEEE Transactions on Medical Imaging, 40 0 (10): 0 2575--2588, 2021

    2021

  9. [17]

    Attentional mechanisms for socially interactive robots--a survey

    Joao Filipe Ferreira and Jorge Dias. Attentional mechanisms for socially interactive robots--a survey. IEEE Transactions on Autonomous Mental Development, 6 0 (2): 0 110--125, 2014

    2014

  10. [18]

    Balanced-mixup for highly imbalanced medical image classification

    Adrian Galdran, Gustavo Carneiro, and Miguel A Gonz \'a lez Ballester. Balanced-mixup for highly imbalanced medical image classification. In Medical Image Computing and Computer Assisted Intervention--MICCAI 2021: 24th International Conference, Strasbourg, France, September 27...

    2021

  11. [19]

    Long-tailed classification of thorax diseases on chest x-ray: A new benchmark study

    Gregory Holste, Song Wang, Ziyu Jiang, Thomas C Shen, George Shih, Ronald M Summers, Yifan Peng, and Zhangyang Wang. Long-tailed classification of thorax diseases on chest x-ray: A new benchmark study. In MICCAI Workshop on Data Augmentation, Labelling, and Imperfections, page...

    2022

  12. [20]

    Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspective

    Muhammad Abdullah Jamal, Matthew Brown, Ming-Hsuan Yang, Liqiang Wang, and Boqing Gong. Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogniti...

    2020

  13. [21]

    Eye-gaze driven surgical workflow segmentation

    Adam James, D Vieira, Benny Lo, Ara Darzi, and G Z Yang. Eye-gaze driven surgical workflow segmentation. In Medical Image Computing and Computer-Assisted Intervention--MICCAI 2007: 10th International Conference, Brisbane, Australia, October 29-November 2, 2007, Proceedings, Pa...

    2007

  14. [22]

    Visual saliency detection by integrating spatial position prior of object with background cues

    Muwei Jian, Jing Wang, Hui Yu, Guodong Wang, Xianjing Meng, Lu Yang, Junyu Dong, and Yilong Yin. Visual saliency detection by integrating spatial position prior of object with background cues. Expert Systems with Applications, 168: 0 114219, 2021

    2021

  15. [23]

    Relational subsets knowledge distillation for long-tailed retinal diseases recognition

    Lie Ju, Xin Wang, Lin Wang, Tongliang Liu, Xin Zhao, Tom Drummond, Dwarikanath Mahapatra, and Zongyuan Ge. Relational subsets knowledge distillation for long-tailed retinal diseases recognition. In Medical Image Computing and Computer Assisted Intervention--MICCAI 2021: 24th I...

    2021

  16. [24]

    Spatial and visual working memory: A mental workspace

    Robert H Logie. Spatial and visual working memory: A mental workspace. In Psychology of learning and motivation, volume 42, pages 37--78. Elsevier, 2003

    2003

  17. [25]

    Eye-gaze-guided vision transformer for rectifying shortcut learning

    Chong Ma, Lin Zhao, Yuzhong Chen, Sheng Wang, Lei Guo, Tuo Zhang, Dinggang Shen, Xi Jiang, and Tianming Liu. Eye-gaze-guided vision transformer for rectifying shortcut learning. IEEE Transactions on Medical Imaging, 2023

    2023

  18. [26]

    Actions in the eye: Dynamic gaze datasets and learnt saliency models for visual recognition

    Stefan Mathe and Cristian Sminchisescu. Actions in the eye: Dynamic gaze datasets and learnt saliency models for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37 0 (7): 0 1408--1424, 2014

    2014

  19. [27]

    Integrating human gaze into attention for egocentric activity recognition

    Kyle Min and Jason J Corso. Integrating human gaze into attention for egocentric activity recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1069--1078, 2021

    2021

  20. [28]

    Robust asymmetric loss for multi-label long-tailed learning

    Wongi Park, Inhyuk Park, Sungeun Kim, and Jongbin Ryu. Robust asymmetric loss for multi-label long-tailed learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2711--2720, 2023

    2023

  21. [29]

    Generalized zero-shot chest x-ray diagnosis through trait-guided multi-view semantic embedding with self-training

    Angshuman Paul, Thomas C Shen, Sungwon Lee, Niranjan Balachandar, Yifan Peng, Zhiyong Lu, and Ronald M Summers. Generalized zero-shot chest x-ray diagnosis through trait-guided multi-view semantic embedding with self-training. IEEE Transactions on Medical Imaging, 40 0 (10): 0...

    2021

  22. [30]

    Balanced meta-softmax for long-tailed visual recognition

    Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems, 33: 0 4175--4186, 2020

    2020

  23. [31]

    Eye tracking validation: Improving radiologist reporting and interpretation

    Mona P Roshan, Jacklyn Garcia, Ana B Cury, Chrisnel Lamy, Frederico Souza, Charif Sidani, and Ricardo C Cury. Eye tracking validation: Improving radiologist reporting and interpretation. European Journal of Radiology, 168: 0 111134, 2023

    2023

  24. [32]

    Svd-clahe boosting and balanced loss function for covid-19 detection from an imbalanced chest x-ray dataset

    Santanu Roy, Mrinal Tyagi, Vibhuti Bansal, and Vikas Jain. Svd-clahe boosting and balanced loss function for covid-19 detection from an imbalanced chest x-ray dataset. Computers in Biology and Medicine, 150: 0 106092, 2022

    2022

  25. [33]

    Transparency of deep neural networks for medical image analysis: A review of interpretability methods

    Zohaib Salahuddin, Henry C Woodruff, Avishek Chatterjee, and Philippe Lambin. Transparency of deep neural networks for medical image analysis: A review of interpretability methods. Computers in biology and medicine, 140: 0 105111, 2022

    2022

  26. [34]

    Action is in the eye of the beholder: Eye-gaze driven model for spatio-temporal action localization

    Nataliya Shapovalova, Michalis Raptis, Leonid Sigal, and Greg Mori. Action is in the eye of the beholder: Eye-gaze driven model for spatio-temporal action localization. Advances in Neural Information Processing Systems, 26, 2013

    2013

  27. [35]

    Gaze-based object segmentation

    Ran Shi, Ngi King Ngan, and Hongliang Li. Gaze-based object segmentation. IEEE Signal Processing Letters, 24 0 (10): 0 1493--1497, 2017

    2017

  28. [36]

    Universal loss reweighting to balance lesion size inequality in 3d medical image segmentation

    Boris Shirokikh, Alexey Shevtsov, Anvar Kurmukov, Alexandra Dalechina, Egor Krivov, Valery Kostjuchenko, Andrey Golanov, and Mikhail Belyaev. Universal loss reweighting to balance lesion size inequality in 3d medical image segmentation. In Medical Image Computing and Computer ...

    2020

  29. [37]

    Gaze locking: passive eye contact detection for human-object interaction

    Brian A Smith, Qi Yin, Steven K Feiner, and Shree K Nayar. Gaze locking: passive eye contact detection for human-object interaction. In Proceedings of the 26th annual ACM symposium on User interface software and technology, pages 271--280, 2013

    2013

  30. [38]

    Understanding cardiology practitioners’ interpretations of electrocardiograms: An eye-tracking study

    Mohammed Tahri Sqalli et al. Understanding cardiology practitioners’ interpretations of electrocardiograms: An eye-tracking study. JMIR Human Factors, 9 0 (1): 0 e34058, 2022

    2022

  31. [39]

    Integrating eye tracking and speech recognition accurately annotates mr brain images for deep learning: proof of principle

    Joseph N Stember, Haydar Celik, David Gutman, Nathaniel Swinburne, Robert Young, Sarah Eskreis-Winkler, Andrei Holodny, Sachin Jambawalikar, Bradford J Wood, Peter D Chang, et al. Integrating eye tracking and speech recognition accurately annotates mr brain images for deep lea...

    2020

  32. [40]

    Digital pathology: the effect of experience on visual search behavior

    Ellhia Sudin, Mitchell Searjeant, George Partridge, Peter Phillips, Louise Hiller, David Snead, Ian Ellis, and Yan Chen. Digital pathology: the effect of experience on visual search behavior. Journal of Medical Imaging, 9 0 (3): 0 035501--035501, 2022

    2022

  33. [41]

    Interpretation of a 12-lead electrocardiogram by medical students: quantitative eye-tracking approach

    Mohammed Tahri Sqalli et al. Interpretation of a 12-lead electrocardiogram by medical students: quantitative eye-tracking approach. JMIR medical education, 7 0 (4): 0 e26675, 2021

    2021

  34. [42]

    Equalization loss for long-tailed object recognition

    Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11662--11671, 2020

    2020

  35. [43]

    Equalization loss v2: A new gradient balance approach for long-tailed object detection

    Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1685--1694, 2021

    2021

  36. [44]

    How visual search relates to visual diagnostic performance: a narrative systematic review of eye-tracking research in radiology

    A Van der Gijp, CJ Ravesloot, H Jarodzka, MF Van der Schaaf, IC Van der Schaaf, Jan PJ van Schaik, and Th J Ten Cate. How visual search relates to visual diagnostic performance: a narrative systematic review of eye-tracking research in radiology. Advances in Health Sciences Ed...

    2017

  37. [45]

    Gazesam: What you see is what you segment

    Bin Wang, Armstrong Aboah, Zheyuan Zhang, and Ulas Bagci. Gazesam: What you see is what you segment. arXiv preprint arXiv:2304.13844, 2023

    2023 arXiv

  38. [46]

    Gazegnn: A gaze-guided graph neural network for chest x-ray classification

    Bin Wang, Hongyi Pan, Armstrong Aboah, Zheyuan Zhang, Elif Keles, Drew Torigian, Baris Turkbey, Elizabeth Krupinski, Jayaram Udupa, and Ulas Bagci. Gazegnn: A gaze-guided graph neural network for chest x-ray classification. In Proceedings of the IEEE/CVF Winter Conference on A...

    2024

  39. [47]

    The devil is in classification: A simple framework for long-tail instance segmentation

    Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is in classification: A simple framework for long-tail instance segmentation. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Pro...

    2020

  40. [48]

    Unified focal loss: Generalising dice and cross entropy-based losses to handle class imbalanced medical image segmentation

    Michael Yeung, Evis Sala, Carola-Bibiane Sch \"o nlieb, and Leonardo Rundo. Unified focal loss: Generalising dice and cross entropy-based losses to handle class imbalanced medical image segmentation. Computerized Medical Imaging and Graphics, 95: 0 102026, 2022

    2022

  41. [49]

    Toward multicenter skin lesion classification using deep neural network with adaptively weighted balance loss

    Guanghui Yue, Peishan Wei, Tianwei Zhou, Qiuping Jiang, Weiqing Yan, and Tianfu Wang. Toward multicenter skin lesion classification using deep neural network with adaptively weighted balance loss. IEEE Transactions on Medical Imaging, 42 0 (1): 0 119--131, 2022

    2022

  42. [50]

    Diagnostic error and bias in the department of radiology: a pictorial essay

    Li Zhang, Xin Wen, Jian-Wei Li, Xu Jiang, Xian-Feng Yang, and Meng Li. Diagnostic error and bias in the department of radiology: a pictorial essay. Insights into Imaging, 14 0 (1): 0 163, 2023 a

    2023

  43. [51]

    Mbnm: multi-branch network based on memory features for long-tailed medical image recognition

    Ruru Zhang, E Haihong, Lifei Yuan, Jiawen He, Hongxing Zhang, Shengjuan Zhang, Yanhui Wang, Meina Song, and Lifei Wang. Mbnm: multi-branch network based on memory features for long-tailed medical image recognition. Computer Methods and Programs in Biomedicine, 212: 0 106448, 2021

    2021

  44. [52]

    Deep long-tailed learning: A survey

    Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023 b

    2023

  45. [53]

    Learning fast sample re-weighting without reward data

    Zizhao Zhang and Tomas Pfister. Learning fast sample re-weighting without reward data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 725--734, 2021

    2021

  46. [54]

    A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlights, and future promises

    S Kevin Zhou, Hayit Greenspan, Christos Davatzikos, James S Duncan, Bram Van Ginneken, Anant Madabhushi, Jerry L Prince, Daniel Rueckert, and Ronald M Summers. A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlight...

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.