Do transformers and CNNs learn different concepts of brain age? DOI Creative Commons
Nys Tjade Siegel, Dagmar Kainmueller, Fatma Deniz

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 9, 2024

Abstract “Predicted brain age” refers to a biomarker of structural health derived from machine learning analysis T1-weighted magnetic resonance (MR) images. A range methods have been used predict age, with convolutional neural networks (CNNs) currently yielding state-of-the-art accuracies. Recent advances in deep introduced transformers, which are conceptually distinct CNNs, and appear set new benchmarks various domains computer vision. However, transformers not yet applied age prediction. Thus, we address two research questions: First, superior CNNs predicting age? Second, do different model architectures learn similar or “concepts age”? We adapted Simple Vision Transformer (sViT) Shifted Window (SwinT) compared both models ResNet50 on 46,381 MR images the UK Biobank. found that SwinT ResNet performed par, while additional training samples will most likely give edge prediction accuracy. identified may characterize (sub-)sets aging effects, representing diverging concepts age. systematically tested whether sViT, focus by examining variations their predictions clinical utility for indicating deviations neurological psychiatric disorders. Reassuringly, did find substantial differences structure between architectures. Based our results, choice architecture does confounding effect studies.

Language: Английский

Comparative Analysis of Brain Age Prediction Using Structural and Diffusion MRIs in Neonates DOI Creative Commons

Zhicong Fang,

Ningning Pan,

Shujuan Liu

et al.

NeuroImage, Journal Year: 2024, Volume and Issue: 299, P. 120815 - 120815

Published: Aug. 25, 2024

Using machine learning techniques to predict brain age from multimodal data has become a crucial biomarker for assessing development. Among various types of imaging data, structural magnetic resonance (sMRI) and diffusion (dMRI) are the most commonly used modalities. sMRI focuses on depicting macrostructural features brain, while dMRI reveals orientation major white matter fibers changes in tissue microstructure. However, their differential capabilities reflecting newborn clinical implications have not been systematically studied. This study aims explore impact prediction. Comparing predictions based T2-weighted(T2w) fractional anisotropy (FA) images, we found mean absolute errors (MAE) predicting infant be similar. Exploratory analysis revealed T2w areas such as cerebral cortex ventricles contribute significantly prediction, whereas FA images highlight regions main tracts. Despite both modalities focusing cortex, they exhibit significant region-wise differences, developmental disparities macro- microstructural aspects cortex. Additionally, examined effects prematurity, gender, hemispherical asymmetry prediction Results showed differences (p<0.05) biases across gender asymmetry, no were observed with images. underscores between age, offering new perspectives studying development aiding more effective assessment tracking

Language: Английский

Citations

4
LSTGINet: Local Attention Spatio-Temporal Graph Inference Network for Age Prediction DOI Creative Commons

Yi Lei,

Xin Wen, Yanrong Hao

et al.

Algorithms, Journal Year: 2025, Volume and Issue: 18(3), P. 138 - 138

Published: March 3, 2025

There is a close correlation between brain aging and age. However, traditional neural networks cannot fully capture the potential age due to limited receptive field. Furthermore, they are more concerned with deep spatial semantics, ignoring fact that effective temporal information can enrich representation of low-level semantics. To address these limitations, local attention spatio-temporal graph inference network (LSTGINet) was developed explore details association aging, taking into account both perspectives. First, multi-scale branches used increase field model simultaneously, achieving perception static correlation. Second, feature graphs reconstructed, large topographies constructed. The node aggregation transfer functions hidden dynamic A new module embedded in component global context establish dependencies interactivity different features, balance differences distribution We use newly designed weighted loss function supervise learning entire prediction framework strengthen process final experimental results show MAE on baseline datasets such as CamCAN NKI 6.33 6.28, respectively, better than current state-of-the-art methods, provides basis for assessing state adults.

Language: Английский

Citations

0
Prompt-guided orthogonal multimodal fusion for cancer survival prediction DOI
Lan Huang,

Shuyu Guo,

Tian Bai

et al.

Information Sciences, Journal Year: 2025, Volume and Issue: unknown, P. 122242 - 122242

Published: April 1, 2025

Language: Английский

Citations

0
Mapping brain development against neurological disorder using contrastive sharing DOI
Muhammad Hassan,

Jieqong Lin,

Ahmed Ameen Fateh

et al.

Expert Systems with Applications, Journal Year: 2024, Volume and Issue: 256, P. 124893 - 124893

Published: Aug. 2, 2024

Language: Английский

Citations

0
MFCA: Collaborative prediction algorithm of brain age based on multimodal fuzzy feature fusion DOI
Weiping Ding, Jing Wang, Jiashuang Huang

et al.

Information Sciences, Journal Year: 2024, Volume and Issue: 687, P. 121376 - 121376

Published: Aug. 19, 2024

Language: Английский

Citations

0
Enhancing perinatal brain maturity estimation using surface deep learning and cross-modal relationship inference technology DOI Creative Commons
Ziyi Yang, Runtao He,

Yucen Sheng

et al.

Research Square (Research Square), Journal Year: 2024, Volume and Issue: unknown

Published: Dec. 23, 2024

Abstract Neonates with marked brain developmental delays are at increased risk of neurodevelopmental disorders. Brain chronological age is a valuable biomarker for assessing abnormal maturation in developing brains; however, accurately estimating birth remains challenging. In this study, we introduce cross-modal relationship inference network (CMRINet) that integrates structural and diffusion magnetic resonance imaging data to improve the accuracy neonatal estimation. The CMRINet employs Transformer encoder relational module capture both long- short-range dependencies multimodal features among cortical parcels. Our model outperformed others predicting age, achieving mean squared error 0.51 absolute 0.55 on test set. By applying trained full-term neonates preterm infants term-equivalent found predicted was significantly lower than suggesting delayed development brains. Furthermore, deviation associated long-term motor infants. These findings highlight effectiveness estimation, potential clinical utility early detection risks during perinatal period.

Language: Английский

Citations

0
Do transformers and CNNs learn different concepts of brain age? DOI Creative Commons
Nys Tjade Siegel, Dagmar Kainmueller, Fatma Deniz

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 9, 2024

Abstract “Predicted brain age” refers to a biomarker of structural health derived from machine learning analysis T1-weighted magnetic resonance (MR) images. A range methods have been used predict age, with convolutional neural networks (CNNs) currently yielding state-of-the-art accuracies. Recent advances in deep introduced transformers, which are conceptually distinct CNNs, and appear set new benchmarks various domains computer vision. However, transformers not yet applied age prediction. Thus, we address two research questions: First, superior CNNs predicting age? Second, do different model architectures learn similar or “concepts age”? We adapted Simple Vision Transformer (sViT) Shifted Window (SwinT) compared both models ResNet50 on 46,381 MR images the UK Biobank. found that SwinT ResNet performed par, while additional training samples will most likely give edge prediction accuracy. identified may characterize (sub-)sets aging effects, representing diverging concepts age. systematically tested whether sViT, focus by examining variations their predictions clinical utility for indicating deviations neurological psychiatric disorders. Reassuringly, did find substantial differences structure between architectures. Based our results, choice architecture does confounding effect studies.

Language: Английский

Citations

0