RapidRad: Accelerating Radiology Report Generation via Distilled Diffusion Models

Diffusion models have recently demonstrated state-of-the-art performance in Radiology Report Generation (RRG). However, their inherently slow, iterative sampling process limits real-time clinical applicability. In this work, we introduce a knowledge distillation framework to accelerate diffusion-based RRG, employing a Latent Consistency Distillation strategy to train a lightweight student model guided by a pre-trained teacher DDPM model. Our approach substantially reduces inference time while preserving the quality and clinical relevance of generated reports, making it a promising solution for efficient, real-time RRG in practical healthcare settings.

کلیدواژه ها:

Radiology Report Generation ، Diffusion Models ، Knowledge Distillation ، Latent Consistency Distillation ، Model Compression ، Chest X-Ray