01
Context
Medical imaging generates massive amounts of data that radiologists must interpret and document. This project aimed to automate the generation of clinically consistent radiology reports from medical images using state-of-the-art deep learning.
02
What I Built
A transformer-based system with a novel relational memory module that maintains context across image sequences. Integrated a pre-trained PubMed BERT model to identify abnormalities and generate medically accurate descriptions. Trained using AWS SageMaker's distributed capabilities.
03
Key Decisions
1Introduced relational memory module for maintaining clinical context
2Integrated pre-trained PubMed BERT for domain-specific language understanding
3Leveraged AWS SageMaker for distributed training on large datasets
4Designed for processing larger datasets of images and reports
04
Challenges
→Ensuring clinical accuracy in generated reports
→Handling large-scale medical imaging datasets
→Balancing model complexity with inference speed
05
Outcomes
✓Achieved superior performance compared to baseline models
✓Successfully processed larger dataset of images and reports
✓Demonstrated practical application of transformers in medical domain
06
Tech Stack
PythonPyTorchAWS SageMakerBERTTransformers