An End-to-End Framework for Ischemic Stroke Lesion Segmentation and Clinical Summary Generation Using Transformer-Enhanced 3D U-Net and LLM
6th International Conference on Machine Learning and Intelligent Systems Engineering, MLISE 2026, Naples, İtalya, 28 - 31 Mayıs 2026, ss.256-261, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/mlise70044.2026.11607644
- Basıldığı Şehir: Naples
- Basıldığı Ülke: İtalya
- Sayfa Sayıları: ss.256-261
- Anahtar Kelimeler: Clinical Reporting, CNN, Deep Learning, LLM, Stroke
- Çukurova Üniversitesi Adresli: Evet
Özet
Accurate and timely identification of ischemic stroke lesions in brain MRI is critical for clinical decision-making, yet manual segmentation remains time-consuming and prone to inter-observer variability. In this work, we propose an end-to-end framework that combines a transformer-enhanced 3D U-Net for volumetric ischemic stroke lesion segmentation with a large language model (LLM) for automatic clinical summary generation. The segmentation model integrates a transformer bottleneck to capture long-range spatial context and achieves a Dice Similarity Coefficient (DSC) of 0.750 on the ISLES 2022 diffusion-weighted imaging (DWI) dataset. Based on the predicted lesion masks, quantitative lesion features are extracted and provided to the LLM to generate radiology-style descriptive summaries. This integrated framework demonstrates the potential of unifying vision and language models to support automated stroke analysis and assist radiological reporting.