ICHOR: A Robust Representation Learning Approach for ASL CBF Maps with Self-Supervised Masked Autoencoders

Xavier Beltran-Urbano; Yiran Li; Xinglin Zeng; Katie R. Jobson; Manuel Taso; Christopher A. Brown; David A. Wolk; Corey T. McMillan

eess.IVMar 5, 2026

ICHOR: A Robust Representation Learning Approach for ASL CBF Maps with Self-Supervised Masked Autoencoders

Authors: Xavier Beltran-Urbano, Yiran Li, Xinglin Zeng, Katie R. Jobson, Manuel Taso, Christopher A. Brown, David A. Wolk, Corey T. McMillan, +5 more

Organizations: University of Pennsylvania, Philadelphia, USA · University of Maryland School of Medicine, Baltimore, USA

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 pretrained 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.

Figures & tables

Figure 1: Overview of the presented study.

Figure 2: Schematic of the ICHOR framework. Stage 1 (top) represents the proposed self-supervised 3D MAE pre-training pipeline. Stage 2 (bottom) illustrates the downstream task adaptation.

Figure 3: Nested 5-fold cross-validation strategy used for evaluation.

Method	CU A $\beta-$ vs CI A $\beta+$					ASL-Quality Prediction
Method	AUC $\uparrow$	Accuracy $\uparrow$	Precision $\uparrow$	Recall $\uparrow$	F1-score $\uparrow$	MSE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$	$R^{2}$ $\uparrow$	PC $\uparrow$
Rand	68.24	57.12	38.5	74.44	50.34	1.21	8.14	10.96	83.99	92.03
MedicalNet [ 1 ]	70.31	68.80	51.07	56.94	51.80	1.31	8.29	11.32	82.54	91.21
BrainSegFounder [ 2 ]	65.90	59.90	40.54	72.5	51.28	1.21	8.25	10.94	83.86	91.98
BrainIAC [ 14 ]	55.97	57.93	37.86	54.72	41.71	1.44	8.72	11.97	80.94	90.30
Ours	78.93	73.54	57.89	63.61	59.37	1.04	7.64	10.15	86.29	93.11

Table 1: Results obtained by adapting the pre-trained encoder to downstream tasks using LoRA. Rand denotes randomly initialized model parameters. Arrows indicate desired direction. Metrics are reported as mean across the five outer folds. (Unit: %)

Figure 4: Reconstruction results for pre-trained models trained with different masking ratios across imaging protocols and scanner vendors.

Metrics	AUC $\uparrow$	Accuracy $\uparrow$	Precision $\uparrow$	Recall $\uparrow$	F1-score $\uparrow$
$\rho=0.25$	76.74	69.54	55.42	65.55	56.94
$\rho=0.5$	78.93	73.54	57.90	63.61	59.37
$\rho=0.75$	71.90	69.44	50.32	61.66	54.67

Table 2: Performance on CU A

\beta-

vs CI A

\beta+

task for different

\rho

values. (Unit: %)

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	N	Range	Mean	SD	Female	Controls	PLD	Dim	Labeling
		Age (years)			Participants (%)		Protocol
ADNI	$966$	55–94	$72.7$	$7.2$	$56.4$	$31.2$	Single	2D	PASL
RSPHA	$422$	56–91	$72.5$	$7.6$	$59.5$	$55.5$	Single	2D	PASL
ALLFTD	$114$	18–84	$52.7$	$15.3$	$49.1$	$48.2$	Multi	3D	PCASL
C-MIND	$100$	0–19	$9.3$	$5.4$	$49.0$	$100.0$	Multi	3D	PCASL
DEPCONN	$187$	20–74	$35.5$	$11.7$	$55.1$	$0.0$	Single	2D	PCASL

Appendix

Table A1: Demographic information of included studies for pre-training.

Demographics and biomarkers
	Diagnosis
	CU (n=103)	MCI (n=35)	AD (n=9)
Age (mean $\pm$ SD)	$71.2\pm 5.7$	$72.5\pm 6.4$	$71.6\pm 7.4$
Female sex, n (%)	68 (66.0)	19 (54.3)	4 (55.6)
Amyloid status (A $\beta$ +/A $\beta$ –)	0/103	35/0	9/0
Acquisition characteristics

Appendix

Table A2: Demographic and acquisition information for the participants included for discriminating CU A

\beta-

vs CI A

\beta+

Demographics
	Diagnosis
	Healthy older adults (n=20)	SVD (n=15)
Age (mean $\pm$ SD)	$72.3\pm 5.1$	$64.3\pm 5.2$
Female sex, n (%)	9 (45.0)	9 (60.0)
Acquisition characteristics
PLD (Single/Multi)	20/0	15/0

Appendix

Table A3: Demographic and acquisition information for the participants included for discriminating HOA vs SVD.

Demographics
	Diagnosis
	AD (n=27)	bvFTD (n=36)
Age (mean $\pm$ SD)	$71.0\pm 7.6$	$64.0\pm 7.4$
Female sex, n (%)	14 (51.9)	12 (33.33)
Acquisition characteristics
PLD (Single/Multi)	27/0	36/0

Appendix

Table A4: Demographic and acquisition information for the participants included for discriminating AD vs bvFTD.

Study	Scanner	N	Age (mean $\pm$ SD)	Female, n (%)	Protocol	PLD
SPRINT		111	$68.7\pm 9.0$	41 (37)	2D/3D PCASL	single
	GE	30	$74.2\pm 7.7$	7 (23)	3D PCASL	single
	Philips	50	$67.3\pm 8.9$	21 (42)	2D PCASL	single
	Siemens	31	$65.5\pm 8.2$	13 (42)	2D PCASL	single
CARDIA		19	$50.5\pm 3.5$	12 (63)	2D PCASL	single
	Siemens	19	$50.5\pm 3.5$	12 (63)	2D PCASL	single

Appendix

Table A5: Demographic and acquisition information for the participants included for predicting ASL quality.

Explore similar work

BrainDINO: A Brain MRI Foundation Model for Generalizable Clinical Representation Learning
Apr 30, 2026Yizhou Wu, Shansong Wang, Yuheng Li +5Medical Imaging Foundation Models Self-Supervised Pre-Training
Department of Radiation Oncology and Winship Cancer Institute, Emory University, Atlanta, GA, USA. · Department of Radiation and Cellular Oncology, The University of Chicago, Chicago, IL, USA. · Department of Biomedical Engineering, Georgia Institute of Technology, Atlanta, GA, USA. +2
3D MRI Image Pretraining via Controllable 2D Slice Navigation Task
May 7, 2026Yu Wang, Qingchao ChenVisual Representation Learning Self-Supervised Pre-Training
Beijing University of Posts and Telecommunications · Peking University
Rhamba: Region-Aware Hybrid Attention-Mamba Framework for Self-Supervised Learning in Resting-State fMRI
May 2, 2026Ruthwik Reddy Doodipala, Pankaj Pandey, Pratheek Eranki +3Self-Supervised Pre-Training Self-Supervised Time Series Representation Learning
1St. Jude Children’s Research Hospital · 2The University of Memphis

Apr 30, 2026J/K move · Enter open · S save

BrainDINO: A Brain MRI Foundation Model for Generalizable Clinical Representation Learning

Yizhou Wu, Shansong Wang, Yuheng Li, Mojtaba Safari+4

Department of Radiation Oncology and Winship Cancer Institute, Emory University, Atlanta, GA, USA. · Department of Radiation and Cellular Oncology, The University of Chicago, Chicago, IL, USA. · Department of Biomedical Engineering, Georgia Institute of Technology, Atlanta, GA, USA.+2

Medical Imaging Foundation Models Self-Supervised Pre-Training Medical Image Representation Learning Brain MRI

Open paper AskPDF arXiv

Brain MRI underpins a wide range of neuroscientific and clinical applications, yet most learning-based methods remain task-specific and require substantial labeled data. Here we show that a single self-supervised representation can generalize across heterogeneous brain MRI endpoints. We trained BrainDINO, a self-distilled foundation model, on approximately 6.6 million unlabeled axial slices from 20 datasets encompassing broad variation in population, disease, and acquisition setting. Using a frozen encoder with lightweight task heads, BrainDINO supported transfer across tumor segmentation, neurodegenerative and neurodevelopmental conditions classification, brain age estimation, post-stroke temporal prediction, molecular status prediction, MRI sequence classification, and survival modeling. Across tasks and supervision regimes, BrainDINO consistently equaled or exceeded natural-image and MRI-specific self-supervised baselines, with particularly strong advantages under label scarcity. Representation analyses further showed anatomically organized and pathology-sensitive feature structure in the absence of task-specific supervision. Our findings indicate that large-scale slice-wise self-supervised learning can yield a unified brain MRI representation that supports diverse neuroimaging tasks without volumetric pretraining or full-network fine-tuning, establishing a scalable foundation for robust and data-efficient brain imaging analysis. Code is available at https://github.com/mclwu22/BrainDINO

Abstract

Explore similar work

BrainDINO: A Brain MRI Foundation Model for Generalizable Clinical Representation Learning

3D MRI Image Pretraining via Controllable 2D Slice Navigation Task

Rhamba: Region-Aware Hybrid Attention-Mamba Framework for Self-Supervised Learning in Resting-State fMRI