Llama3.1 is Now a Radiologist

Researchers at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi have fine-tuned the Llama 3.1 model to develop BiMediX2, a medical AI model. The project codes can be accessed here.

The model excels in both text-based and image-based medical tasks, such as analysing X-rays, CT scans, and MRIs. It is integrated into the messaging platform Telegram, which makes it accessible for telemedicine applications.

“BiMediX2 is the first Arabic-English medical multimodal model, designed to extend inclusive healthcare solutions to over 400 million Arabic speakers,” said Hisham Cholakkal, professor at MBZUAI.

The model, which can interpret medical images and support bilingual interactions, particularly aims to benefit underserved regions in Africa and the Middle East.

“With its bilingual capabilities, this model is a game-changer for telemedicine, enabling better communication and understanding between patients and healthcare providers,” Cholakkal added.

BiMediX2 recently won the Llama Impact Innovation Awards. It was also showcased at the computer expo GITEX in Dubai and the 79th session of the United Nations General Assembly, where it was praised for its potential to bridge healthcare gaps.

MBZUAI researchers credited the open-source Llama ecosystem for accelerating their progress. “The open-source approach allowed us to overcome technical challenges and adapt powerful models like Llama 3.1 for healthcare applications,” Cholakkal further said.

The researchers plan to expand BiMediX2’s capabilities to include more languages and modalities in the future, addressing emerging healthcare challenges and improving access for underserved populations.

Previously, AIM covered how tools like Google’s CT Foundation use AI to process medical imaging data efficiently, allowing researchers to train models with less data and computational resources. 

AI aids in detecting abnormalities, reconstructing images, and enhancing diagnostic insights through methods like generative adversarial networks (GANs) and deep learning tools like UnetU. Despite AI's potential, these technologies remain research-focused and are not yet substitutes for clinical radiologists.