Showing 2 Abstracts.
It hasn’t even been 1 year since chatGPT came out and changed the world. Yet, today, we already have Large Multimodal Models, similar to chatGPT, that can now process images and voice as well. The capabilities of these models have specifically piqued the interest of Radiologists, who have been exploring use cases in both academic research and clinical practice. Researchers at the NIH have developed a GPT that can process radiology reports to process result findings, locally. This means that all the patient health information stays within your network, and you can use GPTs without compromising HIPAA or safety concerns. What’s even greater about the local GPT that the NIH researchers have developed, is that there is no need for costly and time-consuming fine-tuning. The performance of the local GPT developed, just by utilizing some techniques, were on par with the state-of-the-art tool developed by Stanford. The purpose of this educational exhibit is to highlight the capabilities that LMMs have in diagnosing and augmenting the radiologist workflow. I will discuss how there is no need to become a technical expert to ‘fine tune’ models. I will discuss how you can implement LMMs locally to protect health information. I will highlight the advanced LLM techniques that is used to improve the accuracy of outputted results. I will convey the idea that we need to embrace these technologies that are being improved upon at an exponential rate. Ultimately, I will tie this in to how pediatric radiologists can use the techniques I talk about to conduct academic research, and will get people to think about how they can utilize their own local GPTs for various applications in both augmenting the clinical work flow and in academics. Read More
Meeting name: SPR 2024 Annual Meeting & Postgraduate Course , 2024
Authors: Jung Daniel
MRI is crucial to the diagnosis, treatment planning, and monitoring of pediatric posterior fossa tumors; recent advances in machine and deep learning offer the prospect of increased diagnostic accuracy and minimally invasive management strategies. Machine learning and deep learning represent two artificial intelligence (AI) applications of radiomics that show promising results in the identification and classification of pediatric posterior fossa tumors, such as medulloblastoma, diffuse intrinsic pontine glioma, ependymoma, and pilocytic astrocytoma, based on MRI data. Many AI models demonstrate comparable, or in some cases, better, performance than human radiologists. Beyond diagnosis, machine and deep learning algorithms have been developed for tumor histology determination, prognostication, and molecular subtyping, to name just a few examples. In its early stages, however, the technology has several barriers to overcome before mainstream use. Studies of AI algorithms are often conducted with small sample sizes due to the relative rarity of pediatric posterior fossa tumors; many are also conducted at a single institution. Algorithms are susceptible to class imbalance as well, making it more difficult to identify tumors with lower prevalence. Multi-institution studies are important in addressing those limitations, though they have their own cumbersome requirements, such as data harmonization. Deep learning training strategies, such as federated learning, data augmentation, and transfer learning, help to address the problems introduced by working with small data sets. The purpose of this exhibit is to review the methods, results, and limitations identified in studies testing the efficacy of machine and deep learning models for the identification and classification of pediatric posterior fossa tumors and to discuss their potential for use in the clinical setting. Read More
Meeting name: SPR 2024 Annual Meeting & Postgraduate Course , 2024
Authors: Randhawa Hari, Park Brandon, Khanna Praneet, Jung Daniel, Das Ayushman, Ginn Kevin, Mitchell Grace
Keywords: Artificial Intelligence, Deep Learning, Machine Learning