Chest X-Ray Screening and Report Generations
Chest X-rays (CXR) play a crucial role in the diagnosis and treatment of various respiratory and cardiac conditions. Moreover, the need to produce multiple radiology reports daily, coupled with the demand for efficient clinical operations, has created an opportunity for automating this task. Automated report generation can help radiologists and other healthcare professionals to streamline their workflow, reduce the time and effort required for analyzing and interpreting chest X-rays, and improve patient outcomes.
ChestXpert, utilizes advanced transformer-based self-attention approaches to generate more accurate medical reports for CXR (Chest X-ray) images. The encoder of the proposed models uses both CNN and vision transformer for feature extraction which allows extracting both local and global features of the input CXR image, combining the strengths of both models, This combinational approach allows us to leverage the unique strengths of each model to enhance the overall quality and precision of our feature extraction system. Then a transformer-based decoder is used for report generation. Our study concludes that incorporating a combination of a vision transformer and a CNN model for extracting visual features alongside segmented images can enhance the performance of CXR report generation.
Web application https://chestxpert.live/
Datasets: The MIMIC-CXR segmentation dataset is available on physionet.org. The dataset only contains 1141 frontal images selected randomly. Also, they checked the images manually to verify that the lung area is intact and visible.
Segmented dataset: CXLSeg
GitHub code:
System Usability Study for the web application
Project Contributors: Wimukthi Indeewara , Mahela Pradeep , Kasun Rathnayake
Supervisors: Dr. Thanuja Ambegoda, Prof. Dulani Meedeniya