New Saudi AI model to advance radiology diagnosis
KAUST/SDAIA developed MiniGPT-Med outperforms other models
#Saudi #healthcare - King Abdullah University of Science and Technology (KAUST) and Saudi Data and Artificial Intelligence Authority (SDAIA) have released MiniGPT-Med, a new vision-interfaced large language model, designed to provide faster and more efficient radiology diagnosis. The new model was built using Meta’s Llama 2 for the language component and the university’s MiniGPT-v2 as a vision-language alignment backbone.
Built this year under the KAUST-SDAIA partnership programme, MiniGPT-Med demonstrates versatility across various imaging modalities, including X-rays, CT scans, and MRIs covering 14 diseases with results significantly outperforming other state-of-the-art models. The model has been released as open source to promote experimentation and feedback from other researchers and institutions.
SO WHAT? - Radiologists working in many healthcare institutions are overloaded and AI models that aid radiology diagnosis have the potential to fast-track diagnosis and reduce human error. The role played by AI can be life-saving, in particular when it comes to emergencies. However, AI brings its own set of challenges and limitations, including the need for more training data, lack of standardisation in imaging and diagnosis, and a variety of common misinterpretations of medical imagery by AI models. MiniGPT-Med's superior performance in medical image analysis and disease identification advances research for this key medical use case.
Here are a few key details about the development of MiniGPT-Med:
King Abdullah University of Science and Technology (KAUST) and the Saudi Data and Artificial Intelligence Authority (SDAIA) have released MiniGPT-Med, a new vision-interfaced large language model for radiology, built using Meta’s Llama 2 for the language component and MiniGPT-v2 as a vision-language alignment backbone (which was developed at KAUST).
Developed through the KAUST-SDAIA partnership programme, MiniGPT-Med represents several months of intensive collaboration. The model demonstrates proficiency across X-rays, CT scans, and MRIs, covering 14 diseases.
Researchers believe that the development of MiniGPT-Med could advance how AI can be integrated in healthcare, while identifying issues and limitations for further research.
MiniGPT-Med is capable of medical report generation, visual question answering (VQA), and disease identification. It surpasses existing models by 19% in medical report generation accuracy and achieves high-quality outputs in 76% of cases evaluated by radiologists.
MiniGPT-Med significantly outperforms state-of-the-art (SOTA) models across various imaging modalities. The model achieves MIMIC-CXR BERT-Sim performance of 72.0 compared to 53.0 for the SOTA models, and CheXbert-Sim 30.1 compared to 24.9 for SOTA models.
The model architecture is composed of three key components: a visual backbone, a linear projection layer, and an extensive language model.
MiniGPT-Med was trained on more than 670,000 X-rays, CT scans, and MRI images and nearly 6,000 sets of question-answer data, drawn from a variety of sources including MIMIC (Johnson et al., 2019), NLST (The Cancer Imaging Archive, 2023), SLAKE (Medical Visual Question Answering (Med-VQA), 2023), RSNA (Radiological Society of North America, 2018) and RadVQA (OSF, 2023s).
Future plans include diversifying medical datasets and making improvements in understanding complex medical terminology.
MiniGPT-Med is available as open-source code on GitHub to encourage further research and review.
LINKS
Read about other AI models developed in Saudi Arabia:
SDAIA's Arabic LLM now live on watsonx (Middle East AI News)
First LLM trained exclusively on Saudi data sets (Middle East AI News)
World's largest industrial LLM revealed! (Middle East AI News)