The use of machine learning and tractography in hotspot identification for transcranial magnetic stimulation

• Ano de defesa: 2025
• Autor(a) principal: Betioli, Lucas dos Santos
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertacoes da USP; Universidade de São Paulo; Faculdade de Filosofia, Ciências e Letras de Ribeirão Preto
• Palavras-chave em Português: Aprendizado de máquina; Estimulação magnética transcraniana; Tractografia; Machine learning; Tractography; Transcranial magnetic stimulation
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/59/59135/tde-28012026-163238/

Transcranial Magnetic Stimulation (TMS) is a non-invasive technique that uses magnetic pulses to induce currents in specific cortical areas. Accurate coil positioning is essential for mortor area activation, and neuronavigation systems are commonly em ployed to improve placement control. However, identifying stimulation targets is still a time-consuming process that depends heavily on operator expertise. This work investi gates whether deep learning models can generalize across subjects for predicting cortical responsiveness, and compares which type of input information contributes most to the task. Four neural network models were trained on data from five subjects: one using coil coordinates only, one using tractography-derived features only, one using early-fusion of both inputs, and one using late-fusion where features are merged at the final layer. Tractography connectivity features and distal myographic responses were used as data, and model performance was evaluated in both within-subject training and out-of-subject generalization tests using accuracy. Hybrid models combining coil coordinates and fiber information achieved the highest accuracy and F1 in two out of five subjects, outperforming single-input baselines. Nonetheless, substantial inter-subject variability was observed, highlighting the challenges of generalization in this setting. These findings suggest that combining structural connectivity with coil parameters may help automate hotspot local ization and can be extended to other cortical regions and functional domains, potentially reducing the need for exhaustive empirical motor mapping.