NLM IRP Seminar Schedule
UPCOMING SEMINARS
-
May 14, 2024 Stanley Liang
Knowledge-driven Latent Diffusion For COVID-19 Pneumonia Radiology Pattern Synthesis -
May 21, 2024 Ziynet Kesimoglu
TBD -
May 23, 2024 Leslie Ronish
TBD -
May 28, 2024 Harutyun Saakyan
TBD -
May 30, 2024 Deepak Gupta
TBD
RECENT SEMINARS
-
May 9, 2024 Pascal Mutz
The Riboviria protein structurome expands virus protein annotation and highlights protein relations -
May 2, 2024 OPEN
TBD -
April 30, 2024 Wenya Rowe
The conformal central charge of the spin-1/2 XX model derived from long-chain asymptotics -
April 25, 2024 Ermin Hodzic
Condition-Aware Cell Type Deconvolution of Bulk Tissues -
April 16, 2024 Jaya Srivastava
Regulatory plasticity of the human genome
Scheduled Seminars on Feb. 27, 2024
Contact NLM_IRP_Seminar_Scheduling@mail.nih.gov with questions about this seminar.
Abstract:
In medical imaging, leveraging Artificial Intelligence (AI) significantly enhances the precision and efficiency of radiology report generation. Our research introduces two key methodologies that collectively aim to refine the generation and assessment of these reports by integrating AI with the expertise of radiology professionals.
Initially, our approach focuses on improving report preparation by utilizing longitudinal chest X-ray (CXR) data along with historical reports from the MIMIC-CXR dataset. We developed the Longitudinal-MIMIC dataset, a comprehensive collection that incorporates a patient's historical and current visit data, enabling a more informed analysis. This data powers a transformer-based model featuring a cross-attention mechanism and a memory-driven decoder, which pre-fills the 'findings' section of radiology reports by analyzing a patient's past and present CXRs and reports. This technique not only minimizes reporting errors but also enhances the report's accuracy by incorporating extensive patient history.
Moving to the evaluation phase, we integrate the expertise of professional radiologists with the computational efficiency of Large Language Models (LLMs), such as GPT-3.5 and GPT-4. Employing methods like In-Context Instruction Learning (ICIL) and Chain of Thought (CoT) reasoning, our approach aligns AI evaluations with the nuanced judgment of radiology experts. This collaborative model significantly outperforms traditional evaluation metrics, offering a more accurate and detailed assessment of AI-generated reports. The validation of our approach through detailed annotations from radiology professionals sets a new standard for the accurate evaluation of medical reports.
Together, these methodologies represent a synergistic approach to improving radiology report generation and evaluation. By combining longitudinal patient data with expert radiological insight and AI innovation, our work promises to significantly enhance the quality and efficiency of patient care in the field of radiology.