List of Papers Browse by Subject Areas Author List
Abstract
Arterial spin labeling (ASL) perfusion MRI allows direct quantification of regional cerebral blood flow (CBF) without exogenous contrast, enabling noninvasive measurements that can be repeated without constraints imposed by contrast injection. ASL is increasingly acquired in research studies and clinical MRI protocols. Building on successes in structural imaging, recent efforts have implemented deep learning-based methods to improve image quality, enable automated quality control, and derive robust quantitative and predictive biomarkers with ASL-derived CBF. However, progress has been limited by variable image quality, substantial inter-site, vendor and protocol differences, and limited availability of labeled datasets needed to train models that generalize across cohorts. To address these challenges, we introduce ICHOR, a self-supervised pre-training approach for ASL CBF maps that learns transferable representations using 3D masked autoencoders. ICHOR is pre-trained via masked image modeling using a Vision Transformer backbone and can be used as a general-purpose encoder for downstream ASL tasks. For pre-training, we curated one of the largest ASL datasets to date, comprising 11,405 ASL CBF scans from 14 studies spanning multiple sites and acquisition protocols. We evaluated the pre-trained ICHOR encoder on three downstream diagnostic classification tasks and one ASL CBF map quality prediction regression task. Across all evaluations, ICHOR outperformed existing neuroimaging self-supervised pre-training methods adapted to ASL. Pre-trained weights and code will be made publicly available.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5216_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BelXav_ICHOR_MICCAI2026,
author = { Beltran-Urbano, Xavier AND Li, Yiran AND Zeng, Xinglin AND Jobson, Katie R. AND Taso, Manuel AND Brown, Christopher A. AND Wolk, David A. AND McMillan, Corey T. AND Nasrallah, Ilya M. AND Yushkevich, Paul A. AND Wang, Ze AND Detre, John A. AND Dolui, Sudipto},
title = { { ICHOR: A Robust Representation Learning Approach for ASL CBF Maps with Self-supervised Masked Autoencoders } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents ICHOR, a self-supervised pre-training framework tailored for ASL-derived cerebral blood flow (CBF) maps. The work introduces a large-scale multi-site ASL dataset (11,405 scans) and leverages a 3D masked autoencoder with a Vision Transformer backbone to learn transferable representations. The pre-trained encoder is adapted to multiple downstream tasks using LoRA, demonstrating improved performance over existing neuroimaging pre-training models.
- 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) Timely and important problem: The lack of robust representation learning for ASL imaging is a well-recognized gap, and this work directly addresses it. (2) Large-scale dataset contribution: The curated dataset spanning multiple sites and protocols is a significant contribution and supports better generalization. (3) Foundation model perspective: The work goes beyond a single task and proposes a general-purpose encoder for ASL, aligning with current trends in medical foundation models. (4) Comprehensive evaluation: The model is evaluated across multiple downstream tasks, showing consistent improvements. (5) Practical design: The use of LoRA for downstream adaptation is well-motivated and computationally efficient.
- 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 framework largely builds on existing components (MAE, ViT, LoRA) without introducing fundamentally new modeling techniques. (2) The paper does not analyze why MAE-based pretraining is particularly suitable for ASL CBF maps. (3) Some downstream tasks involve limited sample sizes, which may affect the statistical reliability of the reported improvements.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(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 paper makes a strong contribution in terms of dataset scale and the introduction of a dedicated pretraining framework for ASL imaging. While the methodological novelty is limited, the practical impact and potential for broad reuse are substantial. The consistent improvements across multiple tasks further support its effectiveness. Overall, I recommend acceptance.
- 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 #2
- Please describe the contribution of the paper
this work mainly proposed a foundation model or feature extractor for the ASL images, then adapt this model into the following 4 tasks, which show improved performance compared to the previous models.
- 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.
a very large amount of ASL images were collected based on open datasets
- 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 purpose of the proposed work is not clear. If it is foundation model, it should be feasible to other images modalities. The authors seems to build an ASL specific general model or feature extractor, which is also interesting. If so, it should be compared to other ASL models (e.g. ASL image quality control model ASLprep published on nature method journal), other than the general medical images models, which may be not a fair comparison. 2.although a lot of data were collected in this work, the benefit if not clear. Especially, the proposed model was trained on normalized CBF maps. For example, can the model have similar performance with 20% of the data? 3.there is not much model improvement, and it is not clear what is the technical contribution of this work. 4.The downstream tasks is also not typical. For example, using ASL to evaluate A-beta is rarely used in clinic.
- 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 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 major concern is the technical contribution, purpose of the proposed work, and how other researchers can use/replicate it
- 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
The authors employ self-supervised learning on an autoencoder to pretrain a vision transformer-based model designed for ALS MRI images. The SSL task employed was masking using a vision transformer backbone with a goal of reconstructing the original image. The dataset consisted of 11,4005 scans from 14 different sites, one of the largest ASL datasets to date. After initialization, the vision transformer is fine-tuned for downstream tasks using low-rank adaptation on three downstream diagnostic classification tasks and one ASL CBF map quality prediction task. The results demonstrate that the presented transformer outperforms state-of-the-art foundation models.
- 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.
- An ultra-large dataset used to pretrain the model with >11,000 scans.
- There is a compelling demonstration for ASL-specific foundation models.
- The model was able to outperform existing methods on the demonstration downstream tasks with light fine-tuning.
- The pretrained weights will be made publicly available
- 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.
- The two key methodological features of this paper, using a masked autoencoder for SSL and LoRA are well established. The paper does not include significant methodological novelty or advance the underlying SSL pipeline or pretraining.
- Some datasets for downstream tasks are quite small, limiting the ability to accurately assess generalizability, for eg classification between HOA (n=20) and SVD (n=15) or AD (n=27) vs bvFTD (n=36).
- LoRA fine-tuning makes the model highly specific to ASL images and would limit generalizability to other modalities.
- Including a comparison with full finetuning would be beneficial to determine whether the benefits of LoRA are worth the generalizability trade-off.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
Minor comments
- A table segregating performance metrics against ASL scan acquisition parameters would be useful so others could see if their scans fall within parameters for training data.
- The increments for masking could have been coarser than 0.25, 0.5, and 0.75
- 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 makes an important contribution given the scale of the dataset, the pre-trained model, and its performance relative to pre-trained neuroimaging baselines trained on structural MRI. It employs one of the largest ASL datasets to date for, and the paper is technically sound. However, from a methodological standpoint, the paper lacks significant novelty, it relies on existing SSL and adaptation techniques without introducing novel methodological contributions (particularly for MICCAI standards).
- 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
We thank the reviewers for their constructive feedback and for recognizing the importance of ASL-specific representation learning, the value of the large multi-site ASL dataset, and the potential utility of ICHOR as a reusable ASL encoder. Reviewers noted that the work addresses an important gap, uses one of the largest ASL datasets to date, and demonstrates improved performance across downstream tasks.
The reviewers also raised thoughtful critiques and questions, which we have organized into the following themes and address below.
Contribution, positioning, and novelty (R1, R2, R3). We appreciate that reviewers recognized the timeliness and practical impact of developing a representation model for ASL CBF. Reviewers also noted that MAE, ViT, and LoRA are established components rather than new modeling techniques. We acknowledge this point, and emphasize that the contribution is best viewed as an application and resource contribution: integrating these components for ASL CBF at scale, supported by a curated dataset of 11,405 scans from 14 studies and validated across four downstream tasks. ICHOR is intended as an ASL-specific pretrained encoder, rather than a modality-agnostic foundation model. This distinction is central because ASL CBF maps encode physiological perfusion rather than anatomical contrast, and no comparable pretrained backbone currently exists for this modality.
Rationale for MAE-based ASL pretraining (R1). R1 raised an important question about why MAE-style pretraining is suitable for ASL CBF maps. ASL CBF maps are low-SNR, spatially structured physiological images affected by acquisition variability. Reconstructing masked 3D regions from visible context encourages the model to learn distributed perfusion structure and global spatial dependencies, rather than relying only on local intensity patterns. This is particularly relevant for heterogeneous multi-site ASL data, where robust representations must capture shared perfusion structure across variable acquisition conditions.
Evaluation design and downstream task relevance (R1, R2, R3). We appreciate that reviewers recognized the breadth of downstream evaluation, while also noting that some labeled cohorts are small. Limited labeled data is a fundamental constraint in ASL and is a primary motivation for self-supervised pretraining. To reduce instability from small samples, we used nested 5-fold cross-validation with held-out test folds rather than a single split. R2 also raised a question about the clinical relevance of the Aβ task, noting that ASL is not typically used to evaluate amyloid positivity in clinical practice. Our goal is not to propose ASL as a standalone clinical amyloid test. Rather, this task evaluates whether ICHOR captures disease-relevant perfusion patterns that distinguish CU Aβ− controls from cognitively impaired Aβ+ participants, consistent with prior evidence that ASL CBF is sensitive to AD-related neurodegeneration.
Adaptation strategy and additional comparisons (R2, R3). R2 and R3 raised helpful points regarding comparison to existing ASL tools and the choice of LoRA for downstream adaptation. ICHOR is designed as a reusable pretrained encoder, and LoRA was chosen as a parameter-efficient adaptation strategy for limited-label settings and to provide a consistent adaptation protocol across downstream tasks. We agree with R3 that full fine-tuning would be an informative additional comparison and is an important direction for future work. Regarding comparison with ASLPrep and related ASL quality tools (R2), these methods are primarily designed for preprocessing or quality assessment, whereas ICHOR is intended to learn representations that can transfer across multiple task types. Therefore, comparison across diagnostic tasks would not be appropriate. However, a targeted comparison on the ASL quality prediction task would be valuable in future work.
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.
Strengths: The paper makes a strong practical contribution by assembling a very large ASL CBF dataset and introducing an ASL-specific self-supervised masked-autoencoder pretraining framework that shows useful downstream performance with lightweight fine-tuning and promises reusable pretrained weights. Weaknesses: There are some minor concerns of the details of the method, since MAE/ViT/LoRA are established components, ambiguity about whether the goal is a foundation model or ASL-specific feature extractor, small downstream task datasets, and the need for clearer guidance on replication and use by other researchers. The work might be accepted after some minor corrections.