Dual-representation structural MRI classification of psychiatric disorders using deep learning and large language models

NOĞAY H. S., Adeli H.

Psychiatry Research - Neuroimaging, cilt.357, 2026 (SCI-Expanded, Scopus) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 357
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1016/j.pscychresns.2026.112146
  • Dergi Adı: Psychiatry Research - Neuroimaging
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, BIOSIS, EMBASE, MEDLINE, Psycinfo
  • Anahtar Kelimeler: Convolutional neural networks (CNN), Data augmentation (DA), Large language model (LLM), Psychiatric disorders, Segmentation, Structural magnetic resonance imaging (SMRI), Transfer learning (TL)
  • Bursa Uludağ Üniversitesi Adresli: Evet

Özet

Accurate differentiation among psychiatric disorders such as schizophrenia and bipolar disorder remains a significant clinical challenge due to overlapping symptoms and subtle neuroanatomical variations. This study proposes a dual‐representation structural MRI framework in which raw T1-weighted MRI slices and their corresponding color-coded tissue segmentation maps—both derived from the same imaging modality—are analyzed using two independently trained ResNet-18 CNNs. The four diagnostic groups examined include Healthy Controls, Schizophrenia Spectrum, Bipolar Disorder with Psychosis, and Bipolar Disorder without Psychosis. TL and DA techniques were employed to address the limited dataset (N= 103). Following model training, a Large Language Model (LLM) was used as a post-hoc analysis tool to contextualize the CNN outputs and provide interpretative insights into the relative contributions of the two MRI-derived representations. The results indicate that the dual-representation approach improves four-way classification performance and enables systematic comparison of structural information captured by raw versus segmentation-based inputs. These findings highlight the potential of combining deep learning models with LLM-assisted interpretability to support more transparent and informative diagnostic tools in psychiatric neuroimaging.