Cross-Modal Fine-Tuning of 3D Convolutional Foundation Models for ADHD Classification with Low-Rank Adaptation

C.-C. Jay Kuo; Fangxu Xing; Hyun-Hae Cho; Jonghye Woo; Jyun-Ping Kao; Shahar Lazarev; Shinyeong Rho; Taehoon Shin

REVIEW 3 major objections 5 minor 25 references

Cross-Modal Fine-Tuning of 3D Convolutional Foundation Models for ADHD Classification with Low-Rank Adaptation

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read 3D LoRA adapts a CT-pretrained model to ADHD MRI at 71.9% accuracy

desk verdict A plausible 3D-LoRA variant for cross-modal medical-image transfer, but the evaluation protocol is ambiguous enough that the SOTA claim shouldn't be taken at face value. read the letter →

arxiv 2511.06163 v2 pith:UG5NUITZ submitted 2025-11-08 eess.IV cs.CVcs.LGphysics.med-ph

Jyun-Ping Kao , Shinyeong Rho , Shahar Lazarev , Hyun-Hae Cho , Fangxu Xing , Taehoon Shin , C.-C. Jay Kuo , Jonghye Woo This is my paper · ORCID

classification eess.IVcs.CVcs.LGphysics.med-ph

keywords Low-RankAdaptation3Dconvolutionalnetworkscross-modaltransferlearningADHDclassificationdiffusionMRIfoundationmodelparameter-efficientfine-tuningneuroimaging

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

This paper tries to show that a 3D convolutional foundation model pre-trained on CT images can be repurposed to classify ADHD from diffusion MRI by inserting low-rank adapters into every convolutional layer. The adapters factorize each 3D kernel into a pair of rank-4 matrices, so only about 1.64 million parameters are trained while the backbone stays frozen. On a public dataset of 129 youths, one variant reports 71.9% accuracy and another reports 0.716 AUC, both said to beat from-scratch and transfer baselines. If true, the result would make large pre-trained volumetric models practical for small neuroimaging datasets and support the idea that CT-learned features are modality-agnostic enough for MRI-based psychiatric classification.

What carries the argument

The key mechanism is the low-rank factorization of 3D convolutional kernels: a weight tensor W (d_out by d_in·k³) is updated as W' = W + BA, with B (d_out by r) and A (r by d_in·k³) and rank r = 4. In implementation, each 3D kernel is reshaped so the two low-rank matrices act across two spatial dimensions, yielding a 2D low-rank update applied to every 3D filter. An adapter is inserted in parallel with each convolutional layer of the frozen CT-pretrained 3D ResNet-50, and only the adapters plus a two-layer MLP head are trained. This decomposition is what confines training to 1.64 million parameters while preserving the pre-trained feature hierarchy.

What would settle it

Compute the five-fold cross-validated mean accuracy and AUC when a single checkpoint is chosen by a fixed rule (e.g., highest mean validation AUC). If the resulting metrics fall to the level of the baselines (0.600 accuracy, 0.711 AUC), the core claim of a LoRA-driven improvement is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-modal transfer from CT to MRI works for ADHD classification when the adaptation is done with a 3D extension of LoRA. Each 3D convolutional filter is reshaped so a low-rank pair of matrices acts across two of its spatial dimensions, producing a 2D update per filter; these updates are added to the frozen CT-pretrained weights. Across five-fold cross-validation, the ACC-optimized variant reaches 0.719 accuracy and the AUC-optimized variant reaches 0.716 AUC, using 1.64 million trainable parameters — over 113 times fewer than fully fine-tuning the foundation model. The paper interprets this as one of the first successful CT-to-MRI adaptations of a medical

Load-bearing premise

The reported performance assumes that picking the best validation checkpoint in each fold gives an unbiased estimate of test performance; with 129 subjects and selection across two metrics, this assumption is fragile.

Editorial extensions

If this is right

A CT-pretrained encoder can be repurposed for MRI-based classification with only 1.64 million trainable parameters, making large pre-trained 3D models accessible for small neuroimaging cohorts.
The 2D low-rank update per 3D kernel is a general recipe for injecting LoRA into any volumetric CNN, not just ResNet-50.
The accuracy-optimized and AUC-optimized variants separate two clinical use cases: thresholded diagnosis versus risk ranking.
The reported performance establishes a new baseline on this public dataset for ADHD classification from fractional anisotropy and mean diffusivity maps.
Full fine-tuning of the foundation model is shown to be unnecessary; freezing the backbone and adding adapters preserves transferable knowledge.

Reading between the lines

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

Because two checkpoints were selected from validation (one for accuracy, one for AUC) without a statistical correction for that search, the reported test-fold means may overstate general performance; a pre-registered single-checkpoint rule would likely give a less optimistic estimate.
The same low-rank factorization could be tested on other 3D tasks such as tumor segmentation or age prediction to see whether cross-modal transfer generalizes beyond ADHD classification.
The near-chance performance of classical ML baselines suggests that the FA/MD maps carry a weak signal; if the LoRA advantage is real, it implies the CT-pretrained representations supply a strong inductive bias worth studying mechanistically.
The approach could be extended to multi-site or longitudinal data to test whether the learned low-rank updates are site-invariant.

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.

Desk Editor's Note

read the letter

Quick take: the paper's actual contribution is a concrete way to apply LoRA to 3D convolutions—factorizing each 3D kernel into 2D low-rank updates—and it shows that a CT-pretrained 3D ResNet-50 can be adapted to MRI-based ADHD classification with only 1.64M trainable parameters. That is a real, if modest, extension of existing LoRA work, and the method section is clear enough to reproduce in principle. The authors also deserve credit for comparing against a reasonable set of baselines and for acknowledging the single-dataset limitation.

Where it gets soft: the empirical claim of state-of-the-art rests on two checkpoints chosen by 'highest validation accuracy' and 'highest validation AUC' after training. The paper never defines a validation split that is disjoint from the 20% test fold in each of the five folds. If the same 20% is used for both checkpoint selection and test reporting, the reported numbers are not unbiased generalization estimates. This isn't necessarily fatal—they may have held out a portion of the training fold for validation—but the ambiguity is load-bearing, and the reader shouldn't have to guess. Also, the gap over baselines is narrow (AUC 0.716 vs 0.711) and there are no confidence intervals or significance tests on 129 subjects. The ACC-optimized variant actually underperforms the best baseline on AUC (0.630), so 'superior performance' is metric-dependent. Finally, no code is released, which makes the ambiguity harder to resolve.

The method itself is coherent and the paper is honestly written. If I were refereeing this for a conference, I'd ask for a clear statement of the validation protocol, per-fold results with error bars, and ideally an external cohort. But the central idea is worth engaging with.

Who this is for: anyone working on parameter-efficient fine-tuning for medical imaging, or on using foundation models across modalities. It's a useful data point, not a definitive benchmark.

Recommendation: send it to peer review—the idea is worth refereeing—but expect the evaluation to be the main issue. A serious referee should focus on the validation protocol and statistical support.

Referee Report

3 major / 5 minor

Summary. The paper proposes a parameter-efficient transfer-learning method for ADHD classification from diffusion MRI. A 3D ResNet-50 foundation model pre-trained on CT images (FMCIB) is frozen, and low-rank adaptation (LoRA) matrices are inserted into all 3D convolutional layers; the adapted model is fine-tuned on FA and MD maps from a public dataset of 129 subjects. The authors report two model variants selected by validation accuracy and validation AUC, respectively, achieving 71.9% accuracy and 0.716 AUC with only 1.64 million trainable parameters, which they claim is state-of-the-art and 113x more parameter-efficient than full fine-tuning.

Significance. If the reported results are valid, the paper makes a useful empirical contribution: it demonstrates that a CT-pretrained 3D convolutional foundation model can be adapted to MRI-based ADHD classification with a small number of trainable parameters, and it provides a concrete LoRA formulation for 3D convolutional layers. The efficiency gain (1.64M trainable parameters vs 185.57M) is substantial and clearly quantified. However, the central empirical claim rests on a five-fold cross-validation protocol whose checkpoint-selection step is not fully specified, and the reported performance gaps are not accompanied by confidence intervals or significance tests. The two variants are each superior on only one metric, and the AUC margin over the strongest baseline is 0.005. The significance of the paper therefore depends on whether the evaluation protocol can be confirmed and statistically supported.

major comments (3)

[§3.2, Table 1] The evaluation protocol is ambiguous in a load-bearing way. Five-fold cross-validation with an 80/20 split is described, and then the authors state that checkpoints were 'later selected' as the one with highest validation accuracy and the one with highest validation AUC. No validation set disjoint from the 20% held-out test fold is defined. If the same 20% split is used both for checkpoint selection and for the reported test metrics, the final numbers are selection results, not unbiased estimates of generalization, and the SOTA claim collapses. The authors must specify exactly how the validation set was constructed, or re-run the evaluation with a nested or inner-validation protocol that never uses the test fold for model selection.
[Table 1, §4] The claim of 'significantly outperforming' is not supported statistically. With n=129 and five folds, no confidence intervals, per-fold results, or significance tests are reported. The AUC-optimized variant improves over the FMCIB+MLP baseline by only 0.005 (0.716 vs 0.711), which is plausibly within noise. Moreover, 'superior performance' is metric-dependent: the ACC-optimized variant has an AUC of 0.630, below several baselines (e.g., 3D ResNet-50 at 0.710), while the AUC-optimized variant has an accuracy of 0.630, below several DL baselines. The paper should report per-fold metrics, confidence intervals, and paired tests, and should temper the SOTA claim accordingly.
[§2.2] The mathematical description of the proposed LoRA for 3D convolutions is underspecified. The adapted weight is written as W' = W + BA with B in R^{dout x r} and A in R^{r x (din*k^3)}, which is a low-rank update on the flattened kernel, not manifestly a '2D low-rank update per 3D kernel.' The authors mention that kernels are 'effectively reshaped' but do not define the reshape or explain how the update is applied efficiently in the forward pass. This matters for reproducibility and for the novelty claim. Please clarify the exact tensor layout and, if the 2D characterization is essential, give the explicit reshaping and convolution operation.

minor comments (5)

[§1] The statement that applying LoRA to 3D convolutional kernels 'has not been previously addressed' is too strong. LoRA variants for 2D convolutional layers exist, and at least some 3D adaptations have been explored. Please soften the novelty claim and cite representative prior work on convolution LoRA.
[§3.2] For a binary classification task, a softmax over two logits is unusual; typically a sigmoid is used. Please clarify the output layer and the exact definition of the score s_i used for the ROC curve.
[Table 1] The table lists FLOPs but not the number of folds per model or the variance. Adding per-fold accuracy/AUC would greatly improve interpretability. Also, the class imbalance (76 ADHD vs 53 HV) suggests that balanced accuracy or precision-recall AUC should be reported alongside accuracy.
[Throughout] There are typographical issues in the title and header ('FOUNDA TION MODELS', 'V olunteers', 'Y oung') and inconsistent punctuation. A careful copyedit is needed.
[§5] The discussion claims the ACC-optimized model is 'well suited for direct diagnostic use' and the AUC-optimized model is 'valuable for risk ranking.' Given the lack of calibration analysis and the small sample, such clinical-practice claims should be presented as speculative rather than as conclusions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical fine-tuning evaluation; the reported ACC/AUC are measured outcomes, not consequences of the LoRA definition.

full rationale

The paper is an empirical transfer-learning evaluation, not a derivation. The central quantities—five-fold cross-validated ACC and AUC in Table 1—are measured from held-out folds, not derived from the method's definition. The LoRA update W' = W + BA (Sec. 2.2) is imported from prior work [7] and used as a training mechanism; it does not by construction determine the reported accuracy or AUC. No equation in the paper equates a prediction with a fitted input, and no fitted parameter is renamed as a prediction. The only self-citation is the authors' prior 2D LoRA detection paper [9], cited in the introduction as motivation for applying LoRA to vision tasks; it is not load-bearing for the ADHD classification result. The evaluation-protocol ambiguity about whether 'validation' checkpoint selection is disjoint from the held-out test fold is a potential correctness/statistical-validity concern, not a circularity of the derivation: if leakage occurred, the reported numbers would be invalid generalization estimates, but they would still not be forced by the method's definition. Therefore, no significant circularity is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new entities are postulated. All free parameters are training hyperparameters or selection criteria, not physics/domain constants. The main hidden premise is CT-to-DTI transferability, which the paper treats as the hypothesis to be demonstrated rather than as an independent input.

free parameters (5)

LoRA rank r = 4
Chosen without ablation; directly controls the expressiveness of every adapter and hence the reported performance.
Learning rates = 1e-4 (LoRA), 1e-5 (head)
Set by hand; no learning-rate sweep is reported, and the balance between adapter and classifier updates affects the final numbers.
Weight decay = 1e-4
Applied to both parameter groups; no sensitivity analysis is given.
Classifier head hyperparameters = 128 hidden units, GELU, dropout 0.5
Arbitrary architecture choices; no ablations are shown.
Checkpoint selection criterion = best validation ACC or best validation AUC
Two 'representative' models are selected from saved checkpoints based on different validation metrics. Reported test-fold means depend on this post-hoc selection.

assumptions (4)

standard math LoRA additive low-rank update W' = W + BA with rank r approximates full-weight fine-tuning
Taken from Hu et al. [7]; used as a premise without independent derivation in this paper.
domain assumption CT-pretrained FMCIB encodes modality-agnostic representations transferable to FA/MD MRI
Central hypothesis stated in Section 1 and Section 5; tested only indirectly through final classification accuracy/AUC.
domain assumption QSIPrep preprocessing and DTI fitting produce comparable FA/MD maps across subjects
Assumed from pipeline [15]; no motion metrics, quality control, or exclusion details are reported in Section 3.1.
ad hoc to paper A 2D low-rank update per 3D kernel is sufficient to adapt 3D convolutional filters
The specific factorization ('across two of its spatial dimensions', Section 2.2) is not derived or compared with a full 3D low-rank parameterization.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cross-Modal Fine-Tuning of 3D Convolutional Foundation Models for ADHD Classification with Low-Rank Adaptation." pith.science (2026). https://pith.science/paper/UG5NUITZ

@misc{pith2026251106163,
  author       = {Pith},
  title        = {Pith review of: Cross-Modal Fine-Tuning of 3D Convolutional Foundation Models for ADHD Classification with Low-Rank Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UG5NUITZ}},
  note         = {Machine review of arXiv:2511.06163}
}

read the original abstract

Early diagnosis of attention-deficit/hyperactivity disorder (ADHD) in children plays a crucial role in improving outcomes in education and mental health. Diagnosing ADHD using neuroimaging data, however, remains challenging due to heterogeneous presentations and overlapping symptoms with other conditions. To address this, we propose a novel parameter-efficient transfer learning approach that adapts a large-scale 3D convolutional foundation model, pre-trained on CT images, to an MRI-based ADHD classification task. Our method introduces Low-Rank Adaptation (LoRA) in 3D by factorizing 3D convolutional kernels into 2D low-rank updates, dramatically reducing trainable parameters while achieving superior performance. In a five-fold cross-validated evaluation on a public diffusion MRI database, our 3D LoRA fine-tuning strategy achieved state-of-the-art results, with one model variant reaching 71.9% accuracy and another attaining an AUC of 0.716. Both variants use only 1.64 million trainable parameters (over 113x fewer than a fully fine-tuned foundation model). Our results represent one of the first successful cross-modal (CT-to-MRI) adaptations of a foundation model in neuroimaging, establishing a new benchmark for ADHD classification while greatly improving efficiency.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 3 linked inside Pith

[1]

Early diag- nosis is crucial for improving educational, social, and mental health outcomes [2]

INTRODUCTION Attention-deficit/hyperactivity disorder (ADHD) is one of the most common neurodevelopmental disorders, affecting ap- proximately 5–10% of children worldwide [1]. Early diag- nosis is crucial for improving educational, social, and mental health outcomes [2]. However, current ADHD diagnosis re- lies heavily on subjective behavioral assessments...
[2]

Foundation Model Fig

METHOD 2.1. Foundation Model Fig. 1. Overview of the proposed ADHD classification pipeline. Input: A 3D tensor with two channels (FA and MD). Backbone: The input is processed by a pre-trained 3D FM [6]. Multilayer Perceptron (MLP): Performs the final binary classification for ADHD and Healthy V olunteer (HV). We leveraged a 3D ResNet-50 based FM for cance...
[3]

Emotion and Devel- opment Branch Phenotyping and DTI

EXPERIMENTS 3.1. Dataset and Preprocessing This study utilized the open source “Emotion and Devel- opment Branch Phenotyping and DTI” [14] dataset, which comprises a transdiagnostic sample of youth participants aged 7 to 20 years. The original cohort was recruited with inclu- sion criteria for individuals with ADHD, Disruptive Mood Dysregulation Disorder ...

arXiv
[4]

RESULTS Our proposed framework fine-tunes a pre-trained FM via our novel LoRA adaptation strategy, achieving substantially better performance than all conventional baselines in ADHD classification. Table 1 details the five-fold cross-validated mean performance of Machine Learning, Deep Learning (training from scratch or with medical-image pretraining), an...
[5]

This cross-modal adaptation achieved new state-of-the-art results for ADHD diagnosis using only two diffusion MRI-derived feature maps, including FA and MD, as input

DISCUSSION AND CONCLUSION In this work, we introduced a novel application of LoRA in 3D convolutional kernels, demonstrating that a highly PEFT strategy can successfully transfer knowledge from a CT- based FM to an MRI-based ADHD classification task. This cross-modal adaptation achieved new state-of-the-art results for ADHD diagnosis using only two diffus...
[6]

achieved robust performance in oncology imaging. We hypothesize that such a high-capacity CT-based model can be repurposed to detect subtle neuroanatomical patterns in a dif- ferent modality and domain despite the modality and clinical context differences. A key challenge lies in the model’s scale: fine-tuning hundreds of millions of parameters on a small...

arXiv 2026
[7]

COMPLIANCE WITH ETHICAL STANDARDS This research study was conducted retrospectively using hu- man subject data made available in open access by [14]
[8]

Xiaofeng Liu for insightful discus- sions

ACKNOWLEDGMENTS The authors thank Dr. Xiaofeng Liu for insightful discus- sions. This work is partly supported by the National Research Foundation of Korea funded by the Ministry of Science and ICT (RS-2024-00338438)

2024

Show all 25 references

[9]

New insights into attention-deficit/hyperactivity disorder using structural neuroimaging,

Philip Shaw and Cara Rabin, “New insights into attention-deficit/hyperactivity disorder using structural neuroimaging,”Current psychiatry reports, vol. 11, no. 5, pp. 393–398, 2009

2009
[10]

Why the diagnosis of attention deficit hyperactivity dis- order matters,

Alaa M Hamed, Aaron J Kauer, and Hanna E Stevens, “Why the diagnosis of attention deficit hyperactivity dis- order matters,”Frontiers in psychiatry, vol. 6, pp. 168, 2015

2015
[11]

Multimodal mr images-based diagnosis of early adolescent attention- deficit/hyperactivity disorder using multiple kernel learning,

Xiaocheng Zhou, Qingmin Lin, Yuanyuan Gui, Zixin Wang, Manhua Liu, and Hui Lu, “Multimodal mr images-based diagnosis of early adolescent attention- deficit/hyperactivity disorder using multiple kernel learning,”Frontiers in Neuroscience, vol. 15, pp. 710133, 2021

2021
[12]

Population level multimodal neu- roimaging correlates of attention-deficit hyperactivity disorder among children,

Huang Lin, Stefan P Haider, Simone Kaltenhauser, Ali Mozayan, Ajay Malhotra, R Todd Constable, Dustin Scheinost, Laura R Ment, Kerstin Konrad, and Seyed- mehdi Payabvash, “Population level multimodal neu- roimaging correlates of attention-deficit hyperactivity disorder among c...

2023
[13]

Machine-learning-based feature selection to identify attention-deficit hyperactivity disorder using whole- brain white matter microstructure: A longitudinal study,

Huey-Ling Chiang, Chi-Shin Wu, Chang-Le Chen, Wen-Yih Isaac Tseng, and Susan Shur-Fen Gau, “Machine-learning-based feature selection to identify attention-deficit hyperactivity disorder using whole- brain white matter microstructure: A longitudinal study,”Asian Journal of Psyc...

2024
[14]

Foundation model for cancer imaging biomarkers,

Suraj Pai, Dennis Bontempi, Ibrahim Hadzic, Vasco Prudente, Mateo Soka ˇc, Tafadzwa L Chaunzwa, Simon Bernatz, Ahmed Hosny, Raymond H Mak, Nicolai J Birkbak, et al., “Foundation model for cancer imaging biomarkers,”Nature machine intelligence, vol. 6, no. 3, pp. 354–367, 2024

2024
[15]

Lora: Low-rank adaptation of large lan- guage models.,

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al., “Lora: Low-rank adaptation of large lan- guage models.,”ICLR, vol. 1, no. 2, pp. 3, 2022

2022
[16]

Customized segment anything model for medical image segmentation,

Kaidong Zhang and Dong Liu, “Customized segment anything model for medical image segmentation,”arXiv preprint arXiv:2304.13785, 2023

2023 arXiv
[17]

Lora-enhanced rt- detr: First low-rank adaptation based detr for real-time full body anatomical structures identification in muscu- loskeletal ultrasound,

Jyun-Ping Kao, Yu-Ching Chung, Hao-Yu Hung, Chun- Ping Chen, and Wen-Shiang Chen, “Lora-enhanced rt- detr: First low-rank adaptation based detr for real-time full body anatomical structures identification in muscu- loskeletal ultrasound,”Computerized Medical Imaging and Graphi...

2025
[18]

3d u-net: learning dense volumetric segmentation from sparse annotation,

¨Ozg¨un C ¸ ic ¸ek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger, “3d u-net: learning dense volumetric segmentation from sparse annotation,” inInternational conference on medical image computing and computer-assisted intervention. Springer, 2016, pp. 424–432

2016
[19]

Models genesis,

Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang, “Models genesis,”Medi- cal image analysis, vol. 67, pp. 101840, 2021

2021
[20]

Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet?,

Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh, “Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet?,” inProceedings of the IEEE con- ference on Computer Vision and Pattern Recognition, 2018, pp. 6546–6555

2018
[21]

Med3d: Transfer learning for 3d medical image analysis,

Sihong Chen, Kai Ma, and Yefeng Zheng, “Med3d: Transfer learning for 3d medical image analysis,”arXiv preprint arXiv:1904.00625, 2019

1904 arXiv
[22]

”emotion and development branch phenotyping and dti (2012- 2017)

Cameron C. McKay, Brooke Scheinberg, Ellie P. Xu, Katharina Kircanski, Daniel S. Pine, Melissa A. Brot- man, Ellen Leibenluft, and Julia O. Linke, “”emotion and development branch phenotyping and dti (2012- 2017)”,” 2024

2012
[23]

Qsiprep: an integrative plat- form for preprocessing and reconstructing diffusion mri data,

Matthew Cieslak, Philip A Cook, Xiaosong He, Fang- Cheng Yeh, Thijs Dhollander, Azeez Adebimpe, Geof- frey K Aguirre, Danielle S Bassett, Richard F Betzel, Josiane Bourque, et al., “Qsiprep: an integrative plat- form for preprocessing and reconstructing diffusion mri data,”Nat...

2021
[24]

An Inte- grated Approach to Correction for Off-Resonance Ef- fects and Subject Movement in Diffusion MR Imaging,

J. L. R. Andersson and S. N. Sotiropoulos, “An Inte- grated Approach to Correction for Off-Resonance Ef- fects and Subject Movement in Diffusion MR Imaging,” NeuroImage, vol. 125, pp. 1063–1078, January 2016

2016
[25]

A Reproducible Evaluation of ANTs Similarity Metric Performance in Brain Image Registration,

B. B. Avants, N. J. Tustison, G. Song, P. A. Cook, A. Klein, and J. C. Gee, “A Reproducible Evaluation of ANTs Similarity Metric Performance in Brain Image Registration,”NeuroImage, vol. 54, no. 3, pp. 2033– 2044, February 2011

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Early diag- nosis is crucial for improving educational, social, and mental health outcomes [2]

INTRODUCTION Attention-deficit/hyperactivity disorder (ADHD) is one of the most common neurodevelopmental disorders, affecting ap- proximately 5–10% of children worldwide [1]. Early diag- nosis is crucial for improving educational, social, and mental health outcomes [2]. However, current ADHD diagnosis re- lies heavily on subjective behavioral assessments...

[2] [2]

Foundation Model Fig

METHOD 2.1. Foundation Model Fig. 1. Overview of the proposed ADHD classification pipeline. Input: A 3D tensor with two channels (FA and MD). Backbone: The input is processed by a pre-trained 3D FM [6]. Multilayer Perceptron (MLP): Performs the final binary classification for ADHD and Healthy V olunteer (HV). We leveraged a 3D ResNet-50 based FM for cance...

[3] [3]

Emotion and Devel- opment Branch Phenotyping and DTI

EXPERIMENTS 3.1. Dataset and Preprocessing This study utilized the open source “Emotion and Devel- opment Branch Phenotyping and DTI” [14] dataset, which comprises a transdiagnostic sample of youth participants aged 7 to 20 years. The original cohort was recruited with inclu- sion criteria for individuals with ADHD, Disruptive Mood Dysregulation Disorder ...

arXiv

[4] [4]

RESULTS Our proposed framework fine-tunes a pre-trained FM via our novel LoRA adaptation strategy, achieving substantially better performance than all conventional baselines in ADHD classification. Table 1 details the five-fold cross-validated mean performance of Machine Learning, Deep Learning (training from scratch or with medical-image pretraining), an...

[5] [5]

This cross-modal adaptation achieved new state-of-the-art results for ADHD diagnosis using only two diffusion MRI-derived feature maps, including FA and MD, as input

DISCUSSION AND CONCLUSION In this work, we introduced a novel application of LoRA in 3D convolutional kernels, demonstrating that a highly PEFT strategy can successfully transfer knowledge from a CT- based FM to an MRI-based ADHD classification task. This cross-modal adaptation achieved new state-of-the-art results for ADHD diagnosis using only two diffus...

[6] [6]

achieved robust performance in oncology imaging. We hypothesize that such a high-capacity CT-based model can be repurposed to detect subtle neuroanatomical patterns in a dif- ferent modality and domain despite the modality and clinical context differences. A key challenge lies in the model’s scale: fine-tuning hundreds of millions of parameters on a small...

arXiv 2026

[7] [7]

COMPLIANCE WITH ETHICAL STANDARDS This research study was conducted retrospectively using hu- man subject data made available in open access by [14]

[8] [8]

Xiaofeng Liu for insightful discus- sions

ACKNOWLEDGMENTS The authors thank Dr. Xiaofeng Liu for insightful discus- sions. This work is partly supported by the National Research Foundation of Korea funded by the Ministry of Science and ICT (RS-2024-00338438)

2024