Enhanced ChatGPT for Digital Pathology Research
The frontier of AI continues to expand in all industries and the biomedical space is no exception. From analyzing data, identifying therapeutic targets, and developing new medicines, the application of AI within the medical research field seems endless. One challenge presently facing biomedical researchers is figuring out how to accurately and effectively leverage AI when retrieving relevant data. Fortunately, Cornell University is showing promising results with a recent study that spotlights a new, enhanced AI.
Researchers from the Department of Pathology and Laboratory Medicine at Weill Cornell Medicine in New York, NY have published a new study that demonstrates the effectiveness of using domain-specific, text-generative AI. Large language models (LLMs) have caught the attention of many with their capabilities to amass crucial knowledge from vast data stores when provided with any given query. Generative pre-trained transformers like ChatGPT are even mainstream at this point and can be easily accessed and used by the public. However, LLMs have also demonstrated drawbacks with this application, especially when attempting to gather and provide information on niche scientific subjects. Issues of responses containing non-relevant answers, inaccurate sources, or even non-existent hallucinations are common for LLMs.
To address this issue, researchers at Cornell have devised a way to combine a private, domain-specific generative AI tool based on GPT-4 Turbo deployed at Dana Farber Cancer Institute (GPT4DFCI) with a retrieval augmented generation (RAG) architecture. Unafraid of acronyms, they have dubbed this new tool GPT4DFCI-RAG and it draws from a custom curated database of 650 digital pathology publications containing the latest methodologies, applications and datasets. Additionally, researchers processed their data to further enhance the integration between their database and user interactions. To test GPT4DFCI-RAG, researchers pitted it against the generic ChatGPT-4 with a comprehensive set of digital pathology queries and the results were encouraging to say the least.
Across the board, GPT4DFCI-RAG’s answers to the tester’s digital pathology queries were more relevant and accurate. One example cited within the study included a query about available histopathology imaging datasets with nuclei annotations. In this example ChatGPT-4 listed datasets such as The Cancer Genome Atlas and The Cancer Imaging Archive, both of which completely lack the desired annotations. Unlike ChatGPT-4, GPT4DFCI-RAG only responded with valid datasets. In general, ChatGPT-4 provided a high rate of answers with hallucinations, a phenomenon where AI will answer with incorrect or misleading results. Comparatively, GPT4DFCI-RAG had no hallucinations in its responses. With GPT4DFCI-RAG’s encouraging results, researchers go on to explain that the potential implications may benefit ongoing study and education in the future.
With accurate and efficient retrieval of medical data on specialized topics, domain-specific AI could allow researchers to streamline access to knowledge, tackle more crucial tasks, and lessen the cognitive burden of keeping up with current information in their field. This is further evidenced by additional domain-specific tools that have also surfaced and are being used more in biomedicine today. Notable examples include DNAGPT, a GPT derivative for DNA analysis, Med-Gemini, and Med-BERT (Bidirectional Encoder Representation with Transformer), a BERT derivative optimized for structured electronic health records data. Along with these, GPT4DFCI-RAG seems like a welcome addition to the pool of AI innovations in the healthcare space with more improvements to come.
Source:
https://www.thelancet.com/journals/landig/article/PIIS2589-7500(24)00114-6/fulltext
