Signal and Data Processing Journal A scientific journal officially licensed by the Commission for Scientific Publications of the (MSRT). Publisher: Research Ceter for Developmen of Technologies
Volume 23, Issue 1 (6-2026)
JSDP 2026, 23(1): 51-60 |
Back to browse issues page
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Motamed S, Eghdami Arabani M. Detection of intelligence from fMRI images using a combination of long short-term memory network (LSTM) and fuzzy logic (FL). JSDP 2026; 23 (1) : 4
URL: http://jsdp.rcisp.ac.ir/article-1-1443-en.html
URL: http://jsdp.rcisp.ac.ir/article-1-1443-en.html
Assistant Professor, Department of Computer Engineering, FSh.C., Islamic Azad University, Fouman, Iran
Abstract: (591 Views)
Intelligence Quotient (IQ) is a widely used indicator of cognitive ability and is commonly assessed through standardized psychometric tests. However, questionnaire-based IQ assessments may be impractical for certain populations (e.g., very young children) and can be influenced by non-neural factors. In recent years, neuroimaging has enabled the investigation of the relationship between brain structure/function and intelligence. Functional Magnetic Resonance Imaging (fMRI), due to its ability to capture brain activity patterns, provides an opportunity for building automated IQ estimation or IQ-level classification systems. Nevertheless, fMRI data are high-dimensional, noisy, and temporally complex, which motivates the use of deep learning models that can learn sequential dependencies. This study proposes a hybrid framework that combines Long Short-Term Memory (LSTM) networks with a fuzzy inference engine to improve robustness under uncertainty and noise. This research uses the ABIDE dataset, which provides resting-state fMRI data along with phenotypic information. The Full-Scale IQ (FIQ) is considered as the main intelligence criterion. Since FIQ is missing for a subset of subjects, it is computed from Verbal IQ (VIQ) and Performance IQ (PIQ) using the linear relation reported in the manuscript (Relation (1)).
The pipeline consists of three main stages:
The pipeline consists of three main stages:
- Preprocessing: Skull stripping and preparation of brain data, followed by standard fMRI preprocessing steps including noise removal, normalization, realignment, smoothing, ROI selection using the AAL atlas, and extraction of ROI-wise temporal BOLD signals.
- Temporal Feature Learning (LSTM): The extracted ROI time-series are fed to a multi-layer LSTM to capture temporal dynamics and long-range dependencies. The learned latent representation is mapped to a normalized score qLSTMq_{text{LSTM}}qLSTM that summarizes the temporal evidence.
- Fuzzy Inference (TSK): To manage uncertainty in brain signals and improve decision stability, a first-order Takagi–Sugeno–Kang (TSK) fuzzy inference system is employed. The fuzzy system uses qLSTMq_{text{LSTM}}qLSTM along with additional normalized indices (e.g., connectivity and structural indicators defined in the manuscript) to infer the final IQ estimate/class. The rule-base (Table (2)) is used to combine evidence through weighted aggregation of membership degrees.
For evaluation, the data are split into training and test sets, and K-fold cross-validation is used. Performance is reported using accuracy and class-wise F1-score (as illustrated in the corresponding figures), and the proposed method is compared against baseline architectures.
The experimental results indicate that the proposed hybrid LSTM–Fuzzy model achieves higher performance than the baseline methods implemented under the same experimental setting on ABIDE. In particular, the proposed approach reaches an accuracy of 83%, demonstrating a noticeable improvement over conventional deep models and hybrid baselines. The learning curves show a decreasing trend of training and testing errors across boosting/training rounds, suggesting stable convergence. In addition, the class-wise F1-score analysis confirms that the proposed method improves classification consistency across IQ-level categories.
The improvements can be attributed to two complementary factors. First, LSTM effectively models the temporal variability of resting-state fMRI signals and captures sequential dependencies across ROI time-series. Second, fuzzy inference provides a principled mechanism to handle ambiguity and noise by translating numerical evidence into interpretable linguistic rules and aggregating them through membership degrees. This hybrid design increases robustness when the neuroimaging data contain uncertainty, which is common in multi-site resting-state datasets. A key practical aspect of this work is the consistent use of FIQ (reported or computed via Relation (1)), which ensures coherent labeling and evaluation across subjects.
Limitations include dataset heterogeneity (scanner/site effects), computational complexity, and sensitivity to preprocessing choices. Future work may incorporate domain adaptation, transfer learning, or more advanced temporal encoders to improve generalization across acquisition sites.
This study presents a hybrid deep learning and fuzzy logic framework for IQ detection/estimation from resting-state fMRI. By combining LSTM-based temporal feature learning with a TSK fuzzy inference engine, the proposed approach improves performance and robustness on the ABIDE dataset. The results suggest that integrating deep sequential modeling with fuzzy reasoning is an effective strategy for fMRI-based cognitive score analysis.
The improvements can be attributed to two complementary factors. First, LSTM effectively models the temporal variability of resting-state fMRI signals and captures sequential dependencies across ROI time-series. Second, fuzzy inference provides a principled mechanism to handle ambiguity and noise by translating numerical evidence into interpretable linguistic rules and aggregating them through membership degrees. This hybrid design increases robustness when the neuroimaging data contain uncertainty, which is common in multi-site resting-state datasets. A key practical aspect of this work is the consistent use of FIQ (reported or computed via Relation (1)), which ensures coherent labeling and evaluation across subjects.
Limitations include dataset heterogeneity (scanner/site effects), computational complexity, and sensitivity to preprocessing choices. Future work may incorporate domain adaptation, transfer learning, or more advanced temporal encoders to improve generalization across acquisition sites.
This study presents a hybrid deep learning and fuzzy logic framework for IQ detection/estimation from resting-state fMRI. By combining LSTM-based temporal feature learning with a TSK fuzzy inference engine, the proposed approach improves performance and robustness on the ABIDE dataset. The results suggest that integrating deep sequential modeling with fuzzy reasoning is an effective strategy for fMRI-based cognitive score analysis.
Article number: 4
Type of Study: Research |
Subject:
Paper
Received: 2024/10/10 | Accepted: 2026/02/8 | Published: 2026/06/21 | ePublished: 2026/06/21
Received: 2024/10/10 | Accepted: 2026/02/8 | Published: 2026/06/21 | ePublished: 2026/06/21
References
1. G. O. Adebayo, R. V. Yampolskiy, "Estimating intelligence quotient using stylometry and machine learning techniques: A review," Big Data Mining and Analytics, vol. 5, no. 3, pp. 163-191, 2022. [DOI:10.26599/BDMA.2022.9020002]
2. P. Thilakavathy, B. Diwan, "Semi-automatic framework for estimating intelligence quotient (IQ) using MRI," Archives of Clinical and Biomedical Research, pp. 2087-2101, 2021.
3. F. Liu, Y. Shi, Z. Chen, "Intelligence quotient test for smart cities in the United States," Journal of Urban Planning and Development, vol. 147, no. 1, p. 04020053, 2021. [DOI:10.1061/(ASCE)UP.1943-5444.0000637]
4. F. Ke, H. Liu, M. Zhou, R. Yang, H.-M. Cao, "Diagnostic biomarker exploration of autistic patients with different ages and verbal intelligence quotients using random forest," IEEE Access, vol. 9, pp. 123861-123872, 2021. [DOI:10.1109/ACCESS.2021.3071118]
5. A. Arya, M. Manuel, "Intelligence quotient classification from human MRI brain images using convolutional neural network," in Proc. 12th Int. Conf. Computational Intelligence and Communication Networks (CICN), IEEE, 2020, pp. 75-80. [DOI:10.1109/CICN49253.2020.9242552]
6. L. Wang, C. Y. Wee, H. I. Suk, X. Tang, D. Shen, "MRI-based intelligence quotient estimation with sparse learning," PLoS One, vol. 10, no. 3, p. e0117295, 2015. [DOI:10.1371/journal.pone.0117295]
7. Y. Li et al., "Brain anatomical network and intelligence," PLoS Computational Biology, vol. 5, no. 5, p. e1000395, 2009. [DOI:10.1371/journal.pcbi.1000395]
8. M. McDaniel, "Big-brained people are smarter: A meta-analysis of brain volume and intelligence," Intelligence, vol. 33, no. 4, pp. 337-346, 2005. [DOI:10.1016/j.intell.2004.11.005]
9. K. Supekar et al., "Neural predictors of individual differences in response to math tutoring," Proc. Natl. Acad. Sci. USA, vol. 110, no. 20, pp. 8230-8235, 2013. [DOI:10.1073/pnas.1222154110]
10. S. Frangou, X. Chitins, S. C. Williams, "Mapping IQ and gray matter density in healthy young people," NeuroImage, vol. 23, no. 3, pp. 800-805, 2004. [DOI:10.1016/j.neuroimage.2004.05.027]
11. "MRI from picture to proton," American Journal of Neuroradiology, vol. 29, no. 6, p. e50, 2008. [DOI:10.3174/ajnr.A0980]
12. M. A. Balafar et al., "Review of brain MRI image segmentation methods," Artificial Intelligence Review, vol. 33, no. 3, pp. 261-274, 2010. [DOI:10.1007/s10462-010-9155-0]
13. Y. Roy et al., "Deep learning-based EEG analysis: A systematic review," Journal of Neural Engineering, vol. 16, no. 5, 2019. [DOI:10.1088/1741-2552/ab260c]
14. S. S. Yadav, S. Jadhav, "Deep CNN-based medical image classification," Journal of Big Data, vol. 6, no. 1, pp. 1-18, 2019. [DOI:10.1186/s40537-019-0276-2]
15. M. A. Hussain et al., "Can deep learning predict human intelligence from structural MRI?," bioRxiv, 2023. [DOI:10.1101/2023.02.24.529924]
16. T. Shen, "Classifying brain using convolutional neural networks," M.Sc. Thesis, University of Manchester, 2020.
17. Y. S. Vang, Y. Cao, X. Xie, "Deep learning and gradient boosting for intelligence prediction," in ABCD-NP Challenge, MICCAI, 2019.
18. A. El Boustani et al., "MRI brain classification using CNN," in AI2SD, Springer, 2020.
19. S. Pölsterl et al., "Prediction of fluid intelligence from MRI," in MICCAI Workshop, 2019.
20. S. Srivastava et al., "Predicting fluid intelligence using ensemble methods," Springer, 2019.
21. Y. Li et al., "Accounting for temporal variability in fMRI improves IQ prediction," Human Brain Mapping, vol. 44, no. 13, pp. 4772-4791, 2023. [DOI:10.1002/hbm.26415]
22. L. Xiao et al., "Alternating diffusion map-based fusion," IEEE Trans. Biomedical Engineering, vol. 66, no. 8, pp. 2140-2151, 2018. [DOI:10.1109/TBME.2018.2884129]
23. S. Zhang et al., "Multi-task learning model for IQ prediction," Environmental Research, 2024.
24. M. A. Hussain et al., "Deep learning on structural MRI," Scientific Reports, vol. 14, 2024. [DOI:10.1038/s41598-024-78157-0]
25. X. Tong et al., "Connectome signatures from resting-state fMRI," Translational Psychiatry, vol. 12, no. 1, 2022. [DOI:10.1038/s41398-022-02134-2]
26. Y. Sun et al., "Voxel-level brain state prediction using Swin Transformer," arXiv preprint, 2025. [DOI:10.1109/JBHI.2025.3613793]
27. N. He, C. Kou, "Prediction of IQ from resting-state fMRI," Int. J. Developmental Neuroscience, vol. 84, no. 7, pp. 779-790, 2024. [DOI:10.1002/jdn.10375]
28. J. A. Thiele et al., "Choosing explanation over performance," PNAS Nexus, vol. 3, no. 12, 2024. [DOI:10.1093/pnasnexus/pgae519]
29. معتمد، سارا؛ عسکری، احسان، «تشخیص اسلحۀ کمری با استفاده از شبکۀ عصبی کانولوشنی سهبعدی»، فصلنامۀ پردازش علائم و دادهها، دورۀ 2، شمارۀ 56، صفحات 79-69، 1402.
29. Motamed, S., Askari, E., "Detection of handgun using 3D convolutional neural network (3DCNNs)," Signal and Data Processing, vol. 2, no. 56, pp. 69-79, 2023. [DOI:10.61186/jsdp.20.2.69]
30. M. Hanik et al., "Predicting cognitive scores with graph neural networks," Brain Imaging and Behavior, 2021. [DOI:10.21203/rs.3.rs-634170/v1]
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |