Lo Justin, Lim Adam, Wagner Matthias W, Ertl-Wagner Birgit, Sussman Dafna
Department of Electrical, Computer and Biomedical Engineering, Faculty of Engineering and Architectural Sciences, Ryerson University, Toronto, ON, Canada.
Institute for Biomedical Engineering, Science and Technology (iBEST), a partnership between St. Michael's Hospital and Ryerson University, Toronto, ON, Canada.
Front Artif Intell. 2022 Mar 18;5:832485. doi: 10.3389/frai.2022.832485. eCollection 2022.
Rapid development in Magnetic Resonance Imaging (MRI) has played a key role in prenatal diagnosis over the last few years. Deep learning (DL) architectures can facilitate the process of anomaly detection and affected-organ classification, making diagnosis more accurate and observer-independent. We propose a novel DL image classification architecture, Fetal Organ Anomaly Classification Network (FOAC-Net), which uses squeeze-and-excitation (SE) and naïve inception (NI) modules to automatically identify anomalies in fetal organs. This architecture can identify normal fetal anatomy, as well as detect anomalies present in the (1) brain, (2) spinal cord, and (3) heart. In this retrospective study, we included fetal 3-dimensional (3D) SSFP sequences of 36 participants. We classified the images on a slice-by-slice basis. FOAC-Net achieved a classification accuracy of 85.06, 85.27, 89.29, and 82.20% when predicting brain anomalies, no anomalies (normal), spinal cord anomalies, and heart anomalies, respectively. In a comparison study, FOAC-Net outperformed other state-of-the-art classification architectures in terms of class-average F1 and accuracy. This work aims to develop a novel classification architecture identifying the affected organs in fetal MRI.
在过去几年中，磁共振成像（MRI）的快速发展在产前诊断中发挥了关键作用。深度学习（DL）架构可以促进异常检测和受影响器官分类的过程，使诊断更加准确且不依赖观察者。我们提出了一种新颖的DL图像分类架构，即胎儿器官异常分类网络（FOAC-Net），它使用挤压激励（SE）和朴素 inception（NI）模块来自动识别胎儿器官中的异常。这种架构可以识别正常的胎儿解剖结构，还能检测出存在于（1）脑、（2）脊髓和（3）心脏中的异常。在这项回顾性研究中，我们纳入了36名参与者的胎儿三维（3D）稳态自由进动（SSFP）序列。我们逐片对图像进行分类。当预测脑异常、无异常（正常）、脊髓异常和心脏异常时，FOAC-Net的分类准确率分别达到了85.06%、85.27%、89.29%和82.20%。在一项比较研究中，FOAC-Net在类平均F1和准确率方面优于其他现有最先进的分类架构。这项工作旨在开发一种新颖的分类架构，用于识别胎儿MRI中的受影响器官。
