List of Papers Browse by Subject Areas Author List
Abstract
Accurate interpretation of volumetric CT requires efficient navigation of 3D image volumes and attention to diagnostically relevant regions. While eye-tracking has been widely studied in 2D medical imaging, its use for expertise assessment in CT settings remains limited. We propose a gaze-informed transformer framework for expertise classification in thoracic CT. Using a DINOv2 backbone, radiologist fixation patterns are integrated into volumetric feature learning through (1) a learnable log-space bias in self-attention and (2) gaze-weighted pooling of patch embeddings. We trained and evaluated our approach on 182 CT reading sessions from five radiologists with varying levels of experience. On a held-out test set, the model achieves an ROC-AUC of 0.91 and F1 score of 0.86, outperforming adapted methods. These findings suggest that incorporating visual search behavior into transformers may support objective, process-based expertise assessment in radiology. Code is available via https://github.com/leiluk1/GazeToSkill.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1535_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/leiluk1/GazeToSkill
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KhaLei_Predicting_MICCAI2026,
author = { Khaertdinova, Leila AND Anikina, Anna AND Mello-Thoms, Claudia AND Ibragimov, Bulat},
title = { { Predicting Radiologist Expertise from 3D Gaze Patterns During CT Interpretation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents one of the first studies to address skill assessment in radiology by analyzing gaze trajectories within a three-dimensional context. Key contributions include:
- End-to-End Framework: An integrated model that processes 3D CT data alongside corresponding 3D visual attention patterns.
- Novel Gaze Injection: Introduction of “log-space bias” and “gaze-weighted pooling” to effectively incorporate eye-tracking information into the model’s learning process.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Novelty in Dimension: The research significantly extends gaze-based skill assessment from 2D into the 3D dimension, filling a gap where few studies handle both 3D CT and 3D attention maps simultaneously.
Innovative Mechanisms: The proposed log-space bias and gaze-weighted pooling represent novel ways to inject gaze information.
Algorithmic Synergy: These two modules function at complementary levels (the attention layer and the pooling stage), creating a powerful synergy that markedly improves model performance.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Dataset Limitations: The expert dataset is very small, involving only three participants. This makes it difficult to determine if the model captures general expert patterns or merely the specific habits of those individuals.
Lack of Diagnostic Link: The paper lacks validation concerning diagnostic accuracy. Expertise in radiology is ultimately measured by outcomes (e. g. , sensitivity/specificity for detecting nodules), but these metrics are absent.
Clinical Weakness: Without linking gaze patterns to actual diagnostic performance, the clinical claims of the paper are weakened.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The methodology is technically sound and highly novel in its transition to 3D gaze analysis. While the small expert cohort and lack of diagnostic validation are notable limitations, the architectural contributions justify a weak acceptance for the insight they provide into gaze-weighted modeling.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes a gaze-informed vision transformer for classifying radiologist expertise from eye-tracking data during thoracic CT interpretation. The method integrates fixation patterns into a DINOv2 backbone via two mechanisms: a learnable log-space attention bias and gaze-weighted output pooling. The approach is evaluated on a dataset of 182 CT reading sessions from five radiologists, achieving ROC-AUC 0.91 and F1 0.86 on a held-out test set.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The way the authors integrate gaze information into the transformer architecture is clean. The log-space attention bias and gaze-weighted output pooling are two complementary mechanisms that inject gaze information at different stages of the model. The ablation study (Table 2) confirms both contribute (AUC 0.91 combined vs. 0.89 or 0.86 individually).
2.The gaze-to-CT coordinate mapping (gaze drift correction, OCR-based metadata extraction, affine alignment, subject-specific Gaussian calibration) is detailed and addresses real practical challenges in working with eye-tracking data in volumetric imaging. This careful design is needed when dealing with gaze data collection for CTs or volumetric data in general.
3.The evaluation is comprehensive. It covers two distinct families of prior work (skill assessment models and scanpath prediction models) with multiple input combinations. While I feel this work could include more transformer-based methods for comparison, because this is the first work in this domain, I think the current table is sufficient as the metrics, for example F1, are relatively high and reliable and sufficient for publication.
4.The backbone comparison (DINOv2 vs. Med3D, SwinUNETR, UniMISS) and the ablation over attention mode and pooling strategy help justify the design decisions.
5.The authors provide a GitHub link for reproducibility, which is positive and welcomed by the community.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The study uses only five radiologists (three experts, two novices) and 40 CT scans, yielding 182 sessions. This raises concerns about generalizability. The expert/novice distinction is binary based on a 7-year threshold, which is a simplification that ignores the spectrum of expertise. With so few readers, the model may be learning radiologist-specific visual habits rather than generalizable expertise patterns. The test set contains only two radiologists (one expert, one novice), so it is unclear whether the model would transfer to unseen readers with different viewing styles. However, I understand that collecting a large number of experts and CTs requires exponential effort, and given that expert classification can be done reliably by machine learning in the literature, I consider this weakness not a major one.
2.The same 40 CT volumes appear across all radiologists. Since the train and test sets are split by radiologist (A, B, D for training; C, E for testing), the model sees the same CT scans during training and testing, just read by different people. The DINOv2 backbone could learn CT-content features that correlate with which radiologists happened to read which cases, rather than purely learning gaze-based expertise signals. The authors do not discuss or control for this.
3.The paper does not clearly establish who would use this system and in what workflow. In other words, who is the intended user of this system? The potential applications (training feedback, CAD system design) are mentioned only briefly in the final paragraph. There is no discussion of what actionable information a binary expert/novice classification provides beyond what simpler gaze statistics (e.g., total dwell time, fixation count, coverage) could offer. For instance, if an expert were to use this system, what would happen next after the model determines that the reader is an expert? A simple example of a use case in the introduction would be greatly helpful for future readers to understand when they would need this system.
4.Table 1 reports single-point metrics on the test set without error bars or confidence intervals. Given the small test set (67 samples), the reported differences between methods may not be statistically significant. The validation AUC is reported with standard deviation for the ablation (Table 2) but not for the main comparison.
5.The paper does not investigate what gaze patterns the model uses to distinguish experts from novices. There is no attention visualization, feature importance analysis, or qualitative case study showing how the model’s predictions relate to known differences in expert vs. novice reading behavior (e.g., systematic scanning vs. scattered fixations). I find this limiting in terms of interpretability and clinical trust. For example, what if the model simply counts the number of fixations and predicts expertise from that alone? A worthwhile experiment would be to replace the proposed attention and pooling modules with a simple input such as the session length or fixation count, to see how much of the performance can be explained by trivial statistics.
6.While the CT scans come from the publicly available LIDC-IDRI dataset, the eye-tracking data collected by the authors is not shared. This limits reproducibility, especially given that the results depend heavily on the specific gaze data collected from these five individuals.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper tackles an underexplored problem of classifying radiologist expertise from 3D gaze patterns during CT interpretation. The proposed method is clean, integrating fixation data into a vision transformer via log-space attention bias and gaze-weighted pooling, and the ablation study validates that both mechanisms contribute. The data preprocessing pipeline is thorough and the evaluation covers multiple baseline families.
However, several weaknesses requires attention. The dataset is very small (five radiologists, 40 CTs), and the test set contains only two readers, making it hard to assess generalizability beyond individual viewing habits. The potential for data leakage through shared CT volumes across train and test splits is not discussed. The paper also lacks interpretability analysis, leaving open the question of whether the model captures meaningful expertise signals or relies on trivial statistics such as fixation count or session length. The clinical motivation is underdeveloped, with no clear articulation of who would use this system and for what purpose. Finally, the main results lack error bars or statistical significance testing, which is concerning given the small test set.
On balance, I lean toward a weak accept. The methodological contribution is sound and the difficulty of collecting this type of data is acknowledged. But the limited scale, missing baselines, and lack of interpretability prevent me from giving a stronger recommendation. A future version with more radiologists, a trivial-statistics baseline, and some analysis of learned patterns would significantly strengthen the work.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes a method to integrate gaze information from 3D CT scans to predict the expertise level of a radiologist. Specifically, it integrates gaze with images by reweighting attention maps. The authors compare their approach with methods that use only gaze as input (without the underlying image), as well as gaze generation models. The proposed method outperforms these baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The use of gaze data in 3D medical imaging is underexplored and has strong potential for improving downstream tasks. The paper is well-motivated and helps bridge the gap between handcrafted metrics, which are task- and data-agnostic.
- Detailed data collection and preprocessing description.
- Leveraging gaze both as attention across layers (with learnable weights) and for pooling across slices is interesting and simple to implement.
- The proposed method outperforms baselines.
- Leverages gaze scanpath prediction models for expertise classification.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1) Novelty:
- Integrating gaze by reweighting attention requires collapsing across the time dimension to generate a heatmap, which discards temporal information such as the number of refixations that might be informative for expertise classification. 2) Heatmap Preprocessing:
- The motivation and effect of applying L1 normalization to the pooled heatmap are unclear, especially since softmax is already used. The authors should clarify this design choice. 3) Writing:
- It is not clear whether the 3D volume is processed slice-by-slice or as a full volume at the beginning of the method section. It would be helpful to briefly mention or introduce that the volume is partially sampled for clarity. 4) Evaluation:
- Given the relatively small dataset, some form of cross-validation would strengthen the evaluation. In particular, since samples from the same radiologist should not appear in both the training and test sets, it would be informative to evaluate performance across different splits of experts and novices.
- An ablation study on the number of sampled slices (K) would be valuable to understand its impact on performance. This is especially important if there is a difference between K = 1 and K > 1, as it would inform the utility of using gaze on 3D volumes versus 2D images. How did the authors arrive at K = 8?
- The effect of using a fixed \alpha should be analyzed, as it might affect performance.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Since the log-space bias is equivalent to reweighting attention, clarifying the motivation for this formulation could improve clarity. Describing it as a log-space bias may introduce some ambiguity early in the paper.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Gaze in 2D medical imaging has been extensively studied. While 3D volumes may contain richer information, they are also significantly more computationally intensive. Leveraging gaze and integrating it into a model that uses both image and gaze as input is a promising direction and provides a foundation for future methods on 3D volumes. The proposed method also outperforms baseline approaches and utilizes models trained on related tasks (e.g., scanpath prediction) in an interesting way for expertise classification. While there are some weaknesses, including the need for stronger ablations and clearer justification of design choices, the overall contribution is meaningful.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Dear Reviewers, Area and Program Chairs,
We thank you for your overall positive opinion and thoughtful feedback. Below we address the main concerns and present additional analyses on existing data:
Data leakage & experimental design: We clarify that no label leakage occurs from shared CT volumes in train/test splits. The target (expertise) is a property of the reader, not the CT. Each CT carries conflicting labels (paired with both expert and novice gaze), forcing the model to learn gaze dynamics conditioned on the image rather than memorize CT features. Confirming this, a DINOv2 model without gaze yields near-random performance (Test AUC=0.5433), proving CT content alone provides no predictive signal. To directly test for CT-content leakage we evaluated within-CT predictions on the test set. For 29 CTs viewed by both reader C (Expert) and E (Novice), we compared the model’s Expert-probability on the two paired sessions. Under the CT-only hypothesis these probabilities must be equal, yielding within-CT AUC=0.5 and both-correct rate=0% by construction. Instead we observe sign agreement of 86.2% (25/29 CTs with prob_C>prob_E; sign-test p=5.2×10^-5, Wilcoxon p=5.1×10^-6), mean prob_C-prob_E=+0.34, and both paired sessions classified correctly for 65.5% of CTs (with no CT misclassified on both sessions). These results are incompatible with CT-content driving the predictions, thus the model must be conditioning on gaze, which is only other per-session input (
test_stats.txtin GitHub repository). Class-stratified bootstrap on test AUC: 0.909, 95% CI [0.83, 0.97], is strictly above chance.Cross-validation & generalizability: Table 2 ablations use 5-fold reader-disjoint CV on the training pool. A full radiologist-holdout splits across all expert/novice combinations requires >70h of additional training of all methods, which is infeasible within the rebuttal period; these evaluation results will be released on GitHub. We agree with the Reviewers that dataset scale is the primary limitation and we are actively collecting more data. For context, previous 3D CT eye-tracking studies relied on only 2 participants (Pham et al., 2025), and we have 5 participants.
Diagnostic accuracy: We thank the Reviewers for pointing out the importance of diagnostic analysis. In our work, we followed common practice and defined expertise by years of experience. It is also important to note that our novices have <1 year of experience, so no novice is at the edge of being classified as expert. We analyzed readings to measure diagnostic performance: experts achieved 0.96 sensitivity/0.98 specificity, novices 0.89/0.58.This will be included in the Dataset section in the camera-ready version.
Trivial-statistics baseline & interpretability: Per session we computed 21 gaze features: (i) fixation timing: count, total dwell, mean/median/std fixation duration, session duration; (ii) spatial distribution: duration-weighted gaze centroid and spread (x_mean, y_mean, x_std, y_std, dispersion), coverage, entropy; (iii) slice navigation: unique slices visited, slice range, transition count, mean/std inter-frame scroll velocity, revisit ratio; (iv) windowing: unique WL,WW pairs and their change counts. Logistic regression and random forest on reader-disjoint splits yield intra-train 5-fold CV AUC=0.96 but drop to test AUC=0.59 on held-out readers. This rules out the model exploiting trivial statistics and shows handcrafted features fail to generalize across radiologists, while our gaze-conditioned representations transfer.
Method-specific concerns: To clarify, L1 normalization turns heatmaps into probability distributions, making the bias term (α * logw) a log-prior and removing per-radiologist scale differences in raw fixation magnitudes. The model input is the full CT volume. K=8 balances volumetric coverage against compute.
Reproducibility: We agree with the Reviewers that reproducibility is important. We will release a sample of the dataset upon publication.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This submission proposes a gaze-informed transformer framework for expertise classification in thoracic CT, integrating radiologists’ fixation patterns into volumetric feature learning via attention biasing and gaze-weighted pooling. Three reviewers (R1, R2, R3) vote for weak accept. The paper is recognized as a novel and promising contribution, particularly in extending gaze-based expertise assessment from 2D to 3D CT settings. However, all reviewers raise important concerns related to dataset scale, evaluation, and clinical grounding. These issues should be carefully addressed in the rebuttal. All reviewers emphasize that the dataset is small (five radiologists, limited CT scans), raising concerns about generalizability. R2 and R3 raise a serious concern about potential leakage due to shared CT volumes across splits. R3 points out some form of cross-validation would strengthen the evaluation.
