This article demonstrates how to use generative AI and large language models (LLMs) to automatically generate clinical impressions from radiology report findings using AWS services.

• Fine-tunes FLAN-T5 XL model on 91,544 radiology reports from MIMIC-CXR dataset
• Uses Amazon SageMaker JumpStart, Studio for model development and deployment
• Generates impression summaries from findings section of radiology reports
• Evaluates using ROUGE metrics showing 38% improvement over pre-trained model
• Deploys both pre-trained and fine-tuned models as SageMaker endpoints
• Creates QuickSight dashboard to compare inference results with ground truth
• Training on ml.p3.16xlarge instance took approximately 11 hours
• Solution reduces radiologist workload and improves report clarity for patients

The solution demonstrates that fine-tuning general-purpose LLMs for domain-specific clinical tasks significantly outperforms pre-trained models, enabling automated radiology report summarization with practical AWS implementation.