| Summary: | Radiologists examine Computerised Tomography (CT) and Magnetic Resonance Imaging (MRI) scans to synthesise radiology reports. The process involves manual scrutiny of the scans and although TTS software is available, dictating them is strenuous. Additionally, the radiology reports produced may be subjective depending on the radiologist’s findings. By leveraging existing computer vision models that offer segmentation results for brain volume, we can utilise them to enhance radiological reports by incorporating objective quantification of lesion volumes of the brain.
We introduce RadMix, a large language model for radiology that automatically generates the finding section of a radiological report alongside volumetric values of the brain. Through fine tuning of our dataset obtained from MIMIC-III alongside brain volumetric data from HCP, RadMix demonstrates the ability to automatically generate the findings section of a radiology report. It displays how the integration of Natural Language Processing and Computer Vision can further enhance the future of radiology. The successful implementation of RadMix can greatly aid radiologists in the daily tasks of generating a report as the process of writing a radiology report can be labour intensive. The development of a simple toolkit built alongside RadMix provides an avenue for medical professionals to have access to such models to cater to their needs. RadMix aims to simplify the process of writing radiology reports by providing an automated generation of the report to aid junior radiologists during their crafting of the radiology reports. Furthermore, RadMix is a novel model that adds in relevant volumetric brain data in the CT brain scan radiology reports which is a feature of RadMix that is not offered currently. The synthesis of our toolkit and RadMix seeks to cultivate the prospects of AI in healthcare and directly benefit the radiologists’ work.
|
|---|