Evidence record · UDINUS · SDGs
- Unit
- Rektorat UDINUS
- Period
- 2025
- Goals
- Record
- 2930a861-3ad2-4aad-a7f8-22ec3faf8898
- Published
- 2026-10-30
Universitas Dian Nuswantoro conducted a study evaluating deep learning models for the automated classification of brain stroke conditions using MRI images. The activity involved Sukmana, Husni Teja; Hasibuan, Zainal Arifin; Rahman, Abdul Wahab Abdul; Bayuaji, Luhur; and Masruroh, Siti Ummi as participants. The study addressed the importance of early and accurate stroke detection for timely intervention and improved patient outcomes.
UDINUS used a curated dataset of brain MRI scans to train and assess three models: a convolutional neural network baseline, ResNet-18; a standard Vision Transformer, ViT-Base; and a modified Vision Transformer architecture. All models were fine-tuned through transfer learning using consistent preprocessing, training configurations, and evaluation protocols to support a fair comparison. The evaluation examined classification performance, including accuracy, area under the curve, class-wise precision, recall, and F1-scores.
The modified ViT achieved the strongest overall performance, with a validation accuracy of 97.2% and an AUC score of 0.9934. ResNet-18 recorded 92.6% accuracy and an AUC of 0.9982, while ViT-Base recorded 78.6% accuracy and an AUC of 0.8278. The modified ViT showed stronger class-wise precision, recall, and F1-scores for detecting stroke cases, indicating greater sensitivity to critical image features. ResNet-18 remained competitive and efficient, particularly in stroke recall, which was 0.99. ViT-Base underperformed, demonstrating the need for architectural and training adaptations when Vision Transformer models are applied to small medical datasets.
UDINUS also used confusion matrices, ROC curves, integrated gradients, and attention maps to examine the robustness and interpretability of the modified ViT model. The ViT-based models required significantly greater computational resources and longer training times than the CNN baseline, which may limit use in real-time or resource-constrained clinical settings. The activity is presented as UDINUS research on transformer-based medical image classification, examining automated brain stroke classification and the balance between diagnostic performance and computational requirements.