Tu Ethan, Burkow Jonathan, Tsai Andy, Junewick Joseph, Perez Francisco A, Otjen Jeffrey, Alessio Adam M
Michigan State University, Medical Imaging and Data Integration Lab, Department of Biomedical Engineering, East Lansing, Michigan, United States.
Boston Children's Hospital, Harvard Medical School, Boston, Massachusetts, United States.
J Med Imaging (Bellingham). 2024 May;11(3):034505. doi: 10.1117/1.JMI.11.3.034505. Epub 2024 Jun 4.
The limited volume of medical training data remains one of the leading challenges for machine learning for diagnostic applications. Object detectors that identify and localize pathologies require training with a large volume of labeled images, which are often expensive and time-consuming to curate. To reduce this challenge, we present a method to support distant supervision of object detectors through generation of synthetic pathology-present labeled images.
Our method employs the previously proposed cyclic generative adversarial network (cycleGAN) with two key innovations: (1) use of "near-pair" pathology-present regions and pathology-absent regions from similar locations in the same subject for training and (2) the addition of a realism metric (Fréchet inception distance) to the generator loss term. We trained and tested this method with 2800 fracture-present and 2800 fracture-absent image patches from 704 unique pediatric chest radiographs. The trained model was then used to generate synthetic pathology-present images with exact knowledge of location (labels) of the pathology. These synthetic images provided an augmented training set for an object detector.
In an observer study, four pediatric radiologists used a five-point Likert scale indicating the likelihood of a real fracture (1 = definitely not a fracture and 5 = definitely a fracture) to grade a set of real fracture-absent, real fracture-present, and synthetic fracture-present images. The real fracture-absent images scored , real fracture-present images , and synthetic fracture-present images . An object detector model (YOLOv5) trained on a mix of 500 real and 500 synthetic radiographs performed with a recall of and an score of . In comparison, when trained on only 500 real radiographs, the recall and score were and , respectively.
Our proposed method generates visually realistic pathology and that provided improved object detector performance for the task of rib fracture detection.
医学训练数据量有限仍然是机器学习在诊断应用中的主要挑战之一。识别和定位病变的目标检测器需要大量带标注图像进行训练，而精心整理这些图像通常既昂贵又耗时。为了减少这一挑战，我们提出了一种通过生成合成的存在病变标注图像来支持目标检测器远程监督的方法。
我们的方法采用了先前提出的循环生成对抗网络（cycleGAN），并进行了两项关键创新：（1）使用同一受试者相似位置的“近对”存在病变区域和不存在病变区域进行训练；（2）在生成器损失项中添加一个真实度度量（弗雷歇 inception 距离）。我们使用来自704张独特儿科胸部X光片的2800个存在骨折图像块和2800个不存在骨折图像块对该方法进行训练和测试。然后，使用训练好的模型生成具有病变确切位置（标注）的合成存在病变图像。这些合成图像为目标检测器提供了一个扩充的训练集。
在一项观察者研究中，四位儿科放射科医生使用五点李克特量表（1 = 肯定不是骨折，5 = 肯定是骨折）对一组真实的不存在骨折、真实的存在骨折和合成的存在骨折图像进行分级。真实的不存在骨折图像得分 ，真实的存在骨折图像得分 ，合成的存在骨折图像得分 。在500张真实和500张合成X光片混合训练的目标检测器模型（YOLOv5）的召回率为 ，F1分数为 。相比之下，仅在500张真实X光片上训练时，召回率和F1分数分别为 和 。
我们提出的方法生成了视觉上逼真的病变，并且在肋骨骨折检测任务中提高了目标检测器的性能。
