Task-based functional MRI (tb-fMRI) is essential for presurgical brain mapping, targeting crucial networks like language and motor but it requires patient participation, coordination, and time. Resting-state fMRI (rs-fMRI) emerges as a promising alternative [2], capable of mapping multiple networks rapidly at rest, beneficial in patients with deficits or under anaesthesia. Despite rs-fMRI producing more bilateral, extensive, and "noisy" activations [1, 16], this study aims to harness rs-fMRI and Generative AI to mimic task-like activations for the Language network. It explores developing AI to interpret rs-fMRI...

This study utilised raw rs-fMRI data and partially pre-processed tb-fMRI data (n=75 language) from the Human Connectome Project [11-13]. The task-based fMRI was processed using FSL Feat as described by Figure 1.[Fig 1]The rs-fMRI was processed using a custom processing pipeline as seen in Figure 2. The processed rs-fMRI was then parcellated using SENSAAS Language Atlas [14] to derive Functional Connectivity between language related areas.[Fig 2]A 3D deep convolutional Generative Adversarial Network (GAN) [3, 4](Figure 3, 4), featuring progressive growth techniques, was applied to the...

Sensitivity and specificity metrics were calculated for validation and test sets with respect to tb-fMRI [14]. The test sets showed an average sensitivity of 72% and specificity of 99% as seen on Table 1. From Table 2, we can observe that the model achieved 92% sensitivity and 99% specificity on the overall dataset whereas the ICA based method [1, 2, 6, 14] achieved 54.1 % sensitivity and 98.9 % specificity.[Table 1][Table 2]Figure 5 represents the rs-fMRI GAN generated Language activations (in orange colour) and tb-fMRI...

The GAN demonstrated promising results, generating tb-fMRI-like activations from rs-fMRI data. The generative model facilitates quick subject-level analysis of resting-state fMRI for language networks, delivering promising results in seconds. The generative pre-training [8, 9] of the discriminator of the model also enables few-shot learning for labelling ICA components [6] and can be extended to analyse other networks.The model, trained on tb-fMRI data from a specific demographic, may exhibit biases and underperform for other age groups or demographics, necessitating training on a more diverse dataset to...

A. M. Kumaar: Nothing to disclose S. Patalasingh: Nothing to disclose R. Agrawal: Nothing to disclose S. Jain: Nothing to disclose S. P kumaran: Nothing to disclose S. R. Kankara: Research/Grant Support: Part of the research project - Exploring the utility of artificial intelligence augmented functional connectomics approach in brain tumours -a prospective cohort study. N. Menon: Research/Grant Support: Part of the research project - Exploring the utility of artificial intelligence augmented functional connectomics approach in brain tumours -a prospective cohort study. S. A. Rao:...

R. Rolinski et al., “Language lateralization from taskâ€Âbased and resting state functional MRI in patients with epilepsy”, [Online]. Available: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7336139/ P. Branco, “Resting-State Functional Magnetic Resonance Imaging for Language Preoperative Planning”, [Online]. Available: https://www.frontiersin.org/articles/10.3389/fnhum.2016.00011/full A. Radford, L. Metz, and S. Chintala, “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks.” arXiv, Jan. 07, 2016. Accessed: Jul. 21, 2023. [Online]. Available: http://arxiv.org/abs/1511.06434 I. J. Goodfellow et al., “Generative Adversarial Networks.” arXiv, Jun. 10, 2014. Accessed: Jul. 21, 2023. [Online]. Available: http://arxiv.org/abs/1406.2661 N. Nandakumar et al., “A...