Lee Sungwon, Chun Keum San, Lee Seungeun, Park Hyemin, Le Tuan Dinh, Jung Joon-Yong
Department of Radiology, Seoul St. Mary's Hospital, College of Medicine, The Catholic University of Korea, Seoul, 06591, Republic of Korea.
Visual Analysis and Learning for Improved Diagnostics (VALID) Lab, College of Medicine, The Catholic University of Korea, Seoul, 06591, Republic of Korea.
Eur Radiol. 2025 Jun;35(6):3418-3428. doi: 10.1007/s00330-024-11232-2. Epub 2024 Dec 3.
This study validates the use of CycleGAN-generated wrist radiographs with digitally removed splints, specifically assessing their impact on fracture visualisation.
We retrospectively collected wrist radiographs from 1748 patients who had imaging before and after splint application at a single institution. The dataset was divided into training (1696 patients, 5353 images) and testing sets (52 patients, 965 images). A CycleGAN-based model was trained to generate splint-free wrist radiographs (generated "splint-less") from the original "splint" images. A pre-trained fracture detection model (YOLO8s) was used to assess fracture detection performance on three image groups: original "splint-less" radiographs, original "splint" radiographs, and generated "splint-less" radiographs. Two radiologists scored the generated images. Subtraction images quantified overall image alterations. Precision, recall, and F1 scores were used to compare fracture detection performance.
CycleGAN effectively generated splint-suppressed radiographs with minimal remaining splint density (< 10% remaining in 97.99%), hardware distortion (< 10% change in 100%), anatomical distortion (< 10% in 99.63%), and fracture lesion changes (< 10% change in 100%). New artefacts were rare (absent in 97.54%). Notably, the fracture detection model achieved higher precision (0.94 vs. 0.92), recall (0.63 vs. 0.5), and F1 score (0.75 vs. 0.65) on the generated "splint-less" radiographs compared to the original "splint" radiographs, approaching the performance on original "splint-less" radiographs (F1 0.71). Furthermore, greater image alterations by CycleGAN correlated with larger improvements in fracture detection.
CycleGAN successfully removed splint densities from wrist radiographs with splints.
Question Can CycleGAN (Generative Adversarial Networks), designed for image-to-image translation, generate synthetic "splint-less" radiographs to improve fracture visualisation in follow-up radiographs? Findings Removal of splint densities from wrist radiographs using Generative Adversarial Networks preserved anatomical structures and improved the performance of a fracture detection model. Clinical relevance Generated splint-less radiographs can enhance the performance of wrist fracture detection in wrist radiographs, benefiting both human clinicians and AI-powered diagnostic tools.
本研究验证了使用循环生成对抗网络（CycleGAN）生成的去除数字夹板的腕部X光片，特别评估了它们对骨折可视化的影响。
我们回顾性收集了来自一家机构的1748例患者在使用夹板前后的腕部X光片。数据集分为训练集（1696例患者，5353张图像）和测试集（52例患者，965张图像）。基于CycleGAN的模型被训练用于从原始的“有夹板”图像生成无夹板的腕部X光片（生成的“无夹板”图像）。一个预训练的骨折检测模型（YOLO8s）被用于评估在三个图像组上的骨折检测性能：原始的“无夹板”X光片、原始的“有夹板”X光片以及生成的“无夹板”X光片。两名放射科医生对生成的图像进行评分。相减图像量化了整体图像变化。使用精确率、召回率和F1分数来比较骨折检测性能。
CycleGAN有效地生成了夹板抑制的X光片，剩余夹板密度最小（97.99%的图像中剩余<10%），硬件失真（100%的图像中变化<10%），解剖结构失真（99.63%的图像中<10%），以及骨折病变变化（100%的图像中变化<10%）。新的伪影很少见（97.54%的图像中没有）。值得注意的是，与原始的“有夹板”X光片相比，骨折检测模型在生成的“无夹板”X光片上实现了更高的精确率（0.94对0.92）、召回率（0.63对0.5）和F1分数（0.75对0.65），接近在原始“无夹板”X光片上的性能（F1 0.71）。此外，CycleGAN引起的更大图像变化与骨折检测的更大改善相关。
CycleGAN成功地从有夹板的腕部X光片中去除了夹板密度。
问题 为图像到图像翻译设计的循环生成对抗网络（CycleGAN）能否生成合成的“无夹板”X光片以改善后续X光片中的骨折可视化？ 发现 使用生成对抗网络从腕部X光片中去除夹板密度可保留解剖结构并提高骨折检测模型的性能。 临床意义 生成的无夹板X光片可提高腕部X光片中腕部骨折检测的性能，对人类临床医生和人工智能驱动的诊断工具均有益。
