List of Papers Browse by Subject Areas Author List
Abstract
Eye tracking technology as a non-contact method has become an innovative solution to Autism Spectrum Disorder (ASD) diagnosis by comparing the fixation data of healthy individuals and those suffering from ASD. Though this strategy seems promising, challenges remain in ASD data collection stage while the collected data often exhibit large intra-variances that seriously disrupt the diagnostic process. From these two problems, it is evident that binary classification methods which rely on data from both the healthy and ASD groups will encounter difficulties during the deployment stage. Therefore, in this paper, we propose a novel perspective that treats ASD as an abnormality detection through the introduction of one-class classification approach, which only requires healthy individuals’ data during training. Moreover, to address the large intra-variance among individuals, we design a novel uncertainty-aware weighting scheme that assigns images different contribution weights based on their effectiveness in distinguishing between ASD and healthy individuals. Additionally, from uncertainty weighting we are able to conduct an image selection process that selects the best-preforming subset of image stimuli for the training process. In comparison to previous research that adopts the binary classification method, our approach achieves comparable results while only using healthy individuals’ data for the training process.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2859_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/chenf0613/OneClassForASD
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CheFan_Oneclass4ASD_MICCAI2026,
author = { Chen, Fan AND Qi, Jiansong AND Liu, Jie AND Guan, Qingji},
title = { { Oneclass4ASD: Autism Spectrum Disorder Screening via One-class Classifier } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors describe a method for classifying Autistic Spectrum Disorder (ASD) for use as a diagnostic marker, by treating it as an anomaly detection task using a One-Class classification approach on fixation patterns produced by healthy individuals during eye-tracking. The clinical motivation for this is that ASD individuals have uncommon, more erratic gaze patterns than healthy individuals who tend to be more consistent in their response (i. e. what they tend to focus on given a particular image). Rather than restrict themselves to a comparison over a known and fixed diagnostic set of images, they treat this as a general problem for arbitrary input images. Additionally, they introduce a weighting scheme which describes the extent to which each image is likely to produce consistent gaze patterns even among healthy individuals, which is used in two ways; one, as a weighting factor within the classification pipeline itself, and two, as a means of ranking the images so as to then only use the more ‘reliable’ ones for training. They show that this approach achieves better results than more traditional approaches treating healthy and ASD as distinct classes in a binary classification setting.
- 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.
While there is no new core methodology introduced as such, this represents a good, novel pipeline consisting of existing components, solving a useful interesting problem in a straightforward manner.
- 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.
While the broader nature of the pipeline as a whole is mostly clear, the explanations provided can be a bit hand-wavy or lacking in detail at times, making the work harder to follow or be able to replicate the work fully (which is particularly important since no code is provided alongside the paper). Additionally, while they provide ablation studies showing the impact of some components compared to benchmarks, they do not provide much personal insight as to why one approach works and the other does not; they simply use these as evidence that “they happened to make good choices” somehow.
- 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 provide sufficient information for 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
-
Please provide clarifying definitions for all technical terms used in the manuscript, e. g. fixation diagrams, gaze patterns, fixation sequence, etc
-
A (very) brief introduction of what one-class classification methods are would be useful early in the manuscript, to orient the reader who has not come across such methods in the past. Given the broad audience of the conference it’s best not to present advanced concepts under the assumption of full prior knowledge.
-
Please go through your methodology section and clarify all parts and concepts that are ambiguous. E. g. how is the VGG used / set up exactly (e. g. exact nature and dimensionality of inputs, preprocessing involved, and exact nature of outputs); how concatenation is performed (i. e. along what axis), what the 1x1 convolution operation is acting on precisely (i. e. the exact dimensionality of the kernel, otherwise someone unfamiliar with the abuse of terminology common to the field might be confused that this is a 2D operation, effectively a multiplication by a scalar); the exact nature of the “ground-truth fixation map” (is this binary? probabilistic? isolated points? lines? blobs?); what is the exact nature of the “feature vector” (what does 1024-d refer to? are we saying the output of the VGG is 32x32 images expressed as a 1D vector? are the different fixation locations equivalent across images? what is the actual nature of the 1024 vector, is it a one-hot encoding / probabilistic, or does it encode locations in some other manner? what does it mean exactly “to extract a feature from a fixation location”? how are individual fixation locations isolated / differentiated from the probabilistic map?)
-
Can you clarify why is the softmax operation needed after S_i is obtained, why not use the S_i directly (which is already in the [0,1] range).
-
Can you provide some more insight as to why you think Resnet-34 / cosine distance do not work as well as your original choices for this particular problem?
-
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This is a useful clinical contribution, providing good rationale and motivation for the work and appropriate methodological approach for attacking the problem. They demonstrate good results and potential for meaningful use in the relevant clinical domain.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I was happy to recommend acceptance of the manuscript before, and this is still my opinion. Additionally, the authors now state clearly that they are happy to release the source code upon acceptance which was my main concern in terms of reproducibility.
Regarding the author’s rebuttal, I think they address all points adequately. However, I clearly I think all the points they have addressed to the reviewers would have been worth actually addressing in the manuscript, rather than only in the rebuttal, but I see no changes in the manuscript itself. I note the authors say they will make changes post-acceptance? It was my impression that the changes should have been made at this stage so that the reviewers could review these changes, so this is causing me some confusion as to what the actual process is. Nevertheless, I am still of the opinion, as before, that the paper merited acceptance even in its original form.
I agree with the authors’ comments to some of the other reviewers who expressed concerns of clinical usefulness, that the contribution does not need to be an all-or-nothing clinical solution, and that the ability to quickly screen someone is still clinically useful, since these tools do not operate in a vacuum, but typically they are employed in the context of an existing clinical suspicion and a pre-test probability for a diagnosis. Even if it is not highly specific as a test for ASD, if highly sensitive, it can be very useful to quickly rule out an ASD differential (similar to how a d-dimer blood test is used in the emergency department during suspected pulmonary emboli and deep vein thrombosis, for example).
Review #2
- Please describe the contribution of the paper
This paper tests the idea that ASD (Austism Spectrum Disorder) is an example of “not neurotypical” eye tracking data.
They propose to use the similarity of fixation patterns among different neurotypical individuals to improve classification - which produces an improvement.
- 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.
Oneclass approach may reduce reliance on difficult to measure cases, and shows better performance than baseline.
- 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.
THere are no major weaknesses to this paper.
- 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
I would suggest using the term neurotypical instead of healthy, given that healthy can mean many things. Furthermore it seems that the paper casts ASD as abnormal - which can be stigmatising.
I also found the description of the choice of fixation and features were vague
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This appears to be a novel contribution that improves upon a baseline measure across most measures
- 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
Review #3
- Please describe the contribution of the paper
The paper proposes a two-stage one-class classification framework, termed Oneclass4ASD, for ASD screening by modeling it as an anomaly detection problem using only healthy subjects’ data. In the first stage, a dual-scale VGG-16 network is used to predict fixation maps of healthy individuals. In the second stage, fixation points are selected from gaze sequences to extract and concatenate visual features, which are further fused using an uncertainty-aware weighting scheme based on fixation map similarity. Finally, the aggregated features are fed into a Deep SVDD module to learn a hypersphere representing normal patterns and detect abnormal deviations.
- 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 visualizations, together with the proposed framework, provide some evidence that the method is feasible and can effectively capture discriminative gaze-related patterns for ASD screening under the one-class setting.
- 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.
Concern 1: Potential inconsistency in the two-stage design. The proposed two-stage framework raises questions regarding the actual interaction between its components. While the first stage is designed to predict fixation maps for healthy individuals, it is not clearly explained how these predictions are utilized in the second stage. In particular, the selection strategy for a fixed number of fixation points (e.g., 12) is not specified, and its connection to the predicted fixation maps remains unclear. Moreover, the extracted “1024-dimensional visual features” appear to be derived directly from the backbone network applied to the original images, rather than from the predicted fixation maps. This suggests that the first-stage model may not play a significant role during inference. Therefore, it would be helpful for the authors to clarify the dependency between the two stages and justify the necessity of this design. Concern 2: Practical limitations in testing flexibility. The proposed method appears to rely on aggregating features over a relatively large set of observed images (e.g., hundreds of stimuli) for each subject to compute uncertainty-aware weights. This raises questions about its practicality in real-world or clinical scenarios, where collecting such extensive observations for each new subject may not always be feasible. It would be helpful for the authors to clarify whether the method can generalize to settings with fewer observations or more flexible testing conditions. Concern 3: While the use of Deep SVDD enables modeling deviations from healthy patterns, it may be inherently limited in distinguishing the underlying causes of such deviations. For instance, individuals with other neurodevelopmental conditions (e.g., ADHD or intellectual disabilities) may also exhibit atypical gaze patterns. It is therefore unclear how the model would differentiate ASD-specific patterns from other forms of deviation, and whether this could lead to potential misclassification. Further discussion or empirical evaluation on this aspect would strengthen the work. Concern 4: Questionable extent of data efficiency gains. The paper emphasizes alleviating the difficulty of collecting ASD data by adopting a one-class learning paradigm. While this may reduce the need for ASD samples during training, it is less clear whether the overall data collection burden is substantially reduced. For example, the testing phase and the construction of subject-level representations still seem to require a considerable amount of eye-tracking data. It would be valuable for the authors to more carefully discuss the trade-offs and practical implications in terms of data efficiency compared to traditional binary classification approaches.
- Please rate the clarity and organization of this paper
Poor
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The recommendation is based on four main concerns: (1) the two-stage framework lacks clear internal consistency, and the dependency between stages appears insufficiently justified; (2) the testing procedure may require extensive per-subject data aggregation, which could limit practical clinical applicability; (3) although the method claims to alleviate ASD data scarcity, it appears that substantial data collection is still required in both feature construction and testing phases, which may weaken the claimed advantage.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The concerns and criticisms raised in my review should be adequately addressed in the final version.
Review #4
- Please describe the contribution of the paper
1) ASD detection by using anomaly detection formulation. Overcomes the challenges of insufficient and inconsistent ASD fixations. 2) Uncertainty-aware strategy to measure the similarity of fixation pattern from different healthy people, identifying high discriminating stimuli and contribution. 3) The proposed method has reasonable results on the considered dataset without ASD cases.
- 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) problem reformulation aligns with clinical situation. Since ASD data is hard to collect, this method can make the solution easier to implement 2) Robust intra-class variations because of uncertainty-aware weighting. This improves the robustness without introducing extra sample points 3) Good performance without complex system design and modulation.
- 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 dataset itself is too small, just 14 ASD and 14 heathy patients. Overfitting and bias are existing. The results are not convinced. 2) One-class assumption may oversimplify all disorder issues to just anomaly detection. Feature learned may lead to false positive issue with other diseases but not ASD.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The issue is the considered dataset is limited. Although anomaly detection seems to be interesting, but the proposed method setups are too idea for experiment. Hard to judge the generalization and stability on large-scale practical implementation, where multiple similar disorders are existing in the population.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The data sufficiency is still major concern, together with other disease possibility
Author Feedback
We thank Meta-Review (MR) and reviewers for valuable comments.
MR, R4Q1, R5Q2: One-class assumption for multiple disorders. Thanks for your insight. The presence of other mental disorders during testing constitutes a typical open-set scenario, while our work primarily aims to provide autism clinicians with a screening tool for individuals visiting outpatient clinics with suspected symptoms. In open-set settings, traditional binary classification (ASD vs. TD) cannot reliably assess other disorders and only crudely separates individuals into ASD or TD. In contrast, our approach can determine whether an individual is neurotypical in this open-set scenario. Further analysis of other disorders would require additional approaches, such as cross-disorder classification. In future, we will test our approach to other disorders for more in-depth studies.
MR, R4Q2: Two-stage design. Thanks. In our two-stage framework, the first stage trains VGG16 to extract gaze-related features using fixation maps from healthy individuals as supervision. These features guide the network to focus on fixation-relevant patterns but are not used to generate fixation maps. For each subject, the 12 fixation points are selected from the first 12 points in their recorded gaze sequence. The 1024-d features for the second stage are extracted from the backbone using the collected fixation coordinates on the input images, highlighting the necessity of the two-stage design.
MR, R4Q3: Testing flexibility. Although our model requires participants to view hundreds of images, including 3s viewing and 1s break. Using our best model, 100 images takes only 6-7 minutes, while the image usage number can be flexibly adjusted. This free-viewing paradigm is noninvasive and easily implemented for patients.
MR, R5Q1: Dataset size. We address potential overfitting by several complementary strategies across both model design and training. (1) We design a two-stage training framework which decouples the feature extraction and classification. The first stage uses an auxiliary task with fixation maps as supervision to guide feature extraction. This proxy loss instead of a direct classification loss reduces the risk of overfitting. (2) With 300 images per participant in Saliency4ASD, over 8,400 fixation maps provide rich features for the second stage, where a small mu and the soft-boundary objective further mitigate overfitting. (3) we adopt a 14-fold Leave-One-Out strategy to reduce bias, with comparable training and test performance further demonstrating the model’s stability and generalization. Nevertheless, we recognize that increasing the number of participants would further enhance model robustness. Since no other public datasets are available, we are collaborating with hospitals to collect more samples to improve the generalization of our approach in clinical settings.
To R1 Q1-Q3: Clarification. Due to limit space, we cannot give a detailed explanation, but we promise to release the source code upon paper acceptance and carefully revise the paper to clarify our method. Q4-Q5: Softmax, Resnet-34 and cosine. Thanks. Due to page limits, we omitted detailed analysis. (1) Although the scores lie in (0, 1), they are not smooth enough, so Softmax is applied to improve performance. (2) VGG16 outperforms ResNet34 because the smaller fixation maps from ResNet-34 over-smooth the points. (3) Euclidean distance outperforms cosine, as fixation maps carry not only directional patterns but also magnitude and spatial density information.
To R3 Q1-Q2: Clarification. Thanks. We will use neuraltypical for better intuition, and revise the paper carefully.
To R4 Q4: Data collection burden To clarify, our work does not require ASD eye-tracking data for training, which addresses the challenge of the scarcity of ASD samples. While the testing phase requires constructing subject-level representations, each image only needs about 4s for viewing, which is the same as in binary-class method.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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 reframes ASD screening from eye-tracking data as an anomaly detection problem, training only on healthy subjects to avoid reliance on scarce ASD data. Two reviewers (R1, R3) vote for accept and two reviewers (R4, R5) vote for weak reject. Significant concerns remain regarding methodological consistency, dataset scale, and practical applicability. The key points to address in the rebuttal are as follows: R5 highlights that the dataset is extremely small (14 ASD and 14 healthy subjects), raising serious concerns about overfitting, bias, and statistical reliability. R5 also questions whether the one-class assumption oversimplifies real-world clinical scenarios, where multiple disorders may exhibit similar gaze abnormalities. R4 notes that the relationship between stages in the framework is unclear and insufficiently justified. R4 raises concerns that the testing procedure may require extensive per-subject data aggregation, which could limit clinical usability.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Three reviewers recommend acceptance (one rising rating after rebuttal). The responses addressed the concerns about clinical usability and two-stage design.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The main merit of the paper is the reformulation of ASD screening as a one-class anomaly detection problem trained only on healthy subjects. This is a meaningful and clinically relevant perspective, since collecting reliable ASD eye-tracking data is difficult, and a sensitive screening tool can still provide value as part of a broader diagnostic workflow. The proposed framework is well motivated and demonstrates promising results despite the challenging low-data setting. Some concerns remain, particularly regarding the limited statistical reliability due to the small dataset size and the unclear applicability to broader neurodevelopmental conditions beyond ASD. Nevertheless, I think that the practical motivation, problem formulation, and encouraging empirical results outweigh these weaknesses. I therefore recommend acceptance. The remaining limitations should be clearly discussed in the final version.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper was initially invited for rebuttal following the primary review period, where reviewers raised several concerns regarding methodological consistency, dataset scale, and practical applicability. This paper got four reviews, of which three were accept in post-rebuttal. After carefully evaluating the authors’ rebuttal, the reviewers’ revised feedback, and re-examining the paper, I conclude that the authors have effectively addressed the key concerns raised, and the work now meets MICCAI’s publication standards
