University of Twente Student Theses
Vision-language model-based Dutch radiology report generation for lumbar spine X-rays
Poel, C.B. van der (2025) Vision-language model-based Dutch radiology report generation for lumbar spine X-rays.
Full text not available from this repository.
| Full Text Status: | Access to this publication is restricted |
| Abstract: | Lower back pain (LBP) is the leading cause of disability and activity limitation. Diagnostic imaging, such as lumbar spine X-rays (LSXs), can help identify the cause of LBP. However, radiologist interpretation of these X-rays can be time-consuming and variable. Artificial intelligence may assist in addressing these challenges. We aimed to develop a vision-language model for automatic Dutch radiology report generation for LSXs, based on the Large Language and Vision Assistant (LLaVA) architecture. We compared original components (CLIP, Vicuna) with the PubMedCLIP vision encoder and the Dutch GEITje language model, assessed different image preprocessing techniques, and evaluated single- versus multi-turn datasets. Performance was validated using BLEU, ROUGE, and METEOR metrics, alongside radiologist assessment of clinical correctness and impact on patient management. The GEITje-PubMedCLIP model, trained on multi-turn data with zero-padded images, achieved the best results (e.g., BLEU-1 0.0830, METEOR 0.1820, ROUGE-L 0.1518). Radiologist evaluation showed only 20% of reports matched radiologist findings, and 40% would lead to incorrect patient management, mainly due to misidentification of vertebral fractures and spondylolisthesis. While the model generates fluent Dutch reports, improvements are needed in clinical accuracy. Future work should explore multiple input images, further language model and vision encoder enhancements, hyperparameter tuning, and advanced prompting. |
| Item Type: | Essay (Master) |
| Clients: | Ziekenhuisgroep Twente, Almelo, Netherlands |
| Faculty: | TNW: Science and Technology |
| Subject: | 44 medicine, 50 technical science in general, 54 computer science |
| Programme: | Technical Medicine MSc (60033) |
| Link to this item: | https://purl.utwente.nl/essays/106260 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page