Project duration:
Dec 2018 - Sep 2019
dgenrr
Automatic generation of dutch radiology reports from x-ray imagery
Project summary:
While medical expert systems were unsuccessful in the 1960, the artificial intelligence approaches developed since 2000 are very successfully applied in health-care achieving near-human performance in some domains[5]. For instance, deep-neural networks for computer vision detect hipfractures in x-ray images with similar accuracy than human radiologists [4] In the radiology department of the hospital group Twente, radiology images are annotated manually by at least two radiologists. The annotations show a similar structure as indicated in figure 1. In order to make the annotation process more efficient, an automatic detection and annotation framework should be developed. Previous work reported an accuracy of 0.97 for fracture detection [3]. For report generation, 90% of the automatically generated English reports appropriately described the location and 98% appropriately described the character of the fracture [4]. While the latter generates English reports, the linguistic similarity of the Dutch and English language indicate, that the same procedure can be successfully applied to generate high-quality radiology reports in Dutch.
The Goal:
The goal of this project is to generate high-quality Dutch radiology reports from x-ray images of the pelvis based on previous work for hip fractures on x-ray images and reports in English.
Project Leader:
Funding:
ZGT