MedViT-MT: A Multi-Task Medical Vision Transformer for Glioblastoma Brain Analysis with Temporal Encoding and Clinical Data Fusion

Glioblastoma Multiforme (GBM) is the most invasive form of primary brain tumor. The average survival time after diagnosis is only 15 months. Accurately and automatically analyzing GBM MRI scans remains a major clinical challenge. It requires the tumor to be localized, a report to be generated in a structured way, and information from multiple data sources to be integrated simultaneously. Existing methods address these problems separately and use models for one task that do not leverage the complementary information available across segmentation, captioning, and grounding. This study introduces MedViT-MT (Medical Vision Transformer Multi-Task), a unified transformer architecture that performs tumor segmentation, molecular-subtype classification, lesion grounding, and patient-conditioned structured caption generation from GBM DSC-MRI sequences. MedViT-MT employs Swin Transformer Tiny as its hierarchical visual transformer backbone for extracting multiscale imaging features, followed by the addition of a clinical fusion module that is based on cross-attention and that directly places tabular clinical features (IDH1, MGMT, KPS, GTR, and GTR) into the visual feature space. It further extracts inter-slice volumetric information from the MRI slice sequence, helping the model infer three-dimensional tumor volume. On the UPENN-GBM dataset, which contains 375 patients with DSC MRI sequences and their clinical records, MedViT-MT reaches 98.31% classification accuracy, 99.37% AUC-ROC, 91% mean Dice score, and 100% few-shot accuracy (1-shot, 5-shot, 10-shot). The model is also able to generalize well to different IDH1 subtypes, and it runs using 16GB RAM on a CPU. The separate experiments to identify what each architectural component delivers show that all of them matter; the statistical analyses indicate the reliability of the model.

@article{Cheema2026,
  title     = {MedViT-MT: A Multi-Task Medical Vision Transformer for Glioblastoma Brain Analysis with Temporal Encoding and Clinical Data Fusion},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {9},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Sana Cheema and Ghulam Gilanie and Muhammad Faheem Mushtaq and Rania Aboalela and Ali Daud},
  doi       = {10.14569/IJACSA.2026.0170977},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170977}
}