Synthetic Pathology Generation with Near-Pair Cyclic GANs for Object Detectors
Ethan Tu, Jonathan Burkow, Jeffrey P. Otjen, Francisco A. Perez, Joe Junewick, Adam Michael Alessio · 2022
In machine learning for medical imaging, object detectors that localize pathology require a large volume of labeled images. This type of data is often expensive and time-consuming to create. To avoid this challenge, we present a method to support distant supervision of object detectors using synthetic pathology-present labeled images. This method seeks to generate diverse focal pathology and insert this pathology into healthy images. This method is novel in its use of pathology-present regions and similar pathology-absent regions of the same image to serve as "near-pair" cases for the training of a cyclic generative adversarial network (cycleGAN). After training, pathology-absent regions of the image serve as inputs to the trained conditional generator to create synthetic pathology-present regions with exact knowledge of location (labels) in the image. We train and test the method with 2800 fracture-present image patches from 1109 unique pediatric chest radiographs. The trained generator was then used to create fracture-present patches, that were reinserted into radiographs using Poisson inpainting. Visual inspection of the generated patches revealed realistic fracture patterns. In a blinded observer study, three pediatric radiologists were presented with 50 side-by-side images of a real and synthetic radiograph and asked to identify the real case. Radiologists achieved an accuracy of 67±17% for the selection of the real vs. synthetic case, where 50% accuracy would reflect equivocal distinction between the two. In 70% of cases, at least one radiologist did not correctly identify the real case, and radiologists unanimously agreed in only 33% of cases. These results suggest that the proposed method creates visually realistic pathology. Our methodology is potentially generalizable to other imaging applications with focal pathology such as tumor-present PET imaging.