Department of Radiology, Boston Medical Center, Boston University School of Medicine, FGH Building, 3rd Floor, 820 Harrison Avenue, Boston, MA, 02118, USA.
Research Center for Cooperative Projects, Hokkaido University Graduate School of Medicine, Sapporo, Hokkaido, Japan.
Eur Radiol. 2021 Jul;31(7):5206-5211. doi: 10.1007/s00330-020-07568-0. Epub 2021 Jan 6.
Diagnosis of otosclerosis on temporal bone CT images is often difficult because the imaging findings are frequently subtle. Our aim was to assess the utility of deep learning analysis in diagnosing otosclerosis on temporal bone CT images.
A total of 198 temporal bone CT images were divided into the training set (n = 140) and the test set (n = 58). The final diagnosis (otosclerosis-positive or otosclerosis-negative) was determined by an experienced senior radiologist who carefully reviewed all 198 temporal bone CT images while correlating with clinical and intraoperative findings. In deep learning analysis, a rectangular target region that includes the area of the fissula ante fenestram was extracted and fed into the deep learning training sessions to create a diagnostic model. Transfer learning was used with the deep learning model architectures of AlexNet, VGGNet, GoogLeNet, and ResNet. The test data set was subsequently analyzed using these models and by another radiologist with 3 years of experience in neuroradiology following completion of a neuroradiology fellowship. The performance of the radiologist and the deep learning models was determined using the senior radiologist's diagnosis as the gold standard.
The diagnostic accuracies were 0.89, 0.72, 0.81, 0.86, and 0.86 for the subspecialty trained radiologist, AlexNet, VGGNet, GoogLeNet, and ResNet, respectively. The performances of VGGNet, GoogLeNet, and ResNet were not significantly different compared to the radiologist. In addition, GoogLeNet and ResNet demonstrated non-inferiority compared to the radiologist.
Deep learning technique may be a useful supportive tool in diagnosing otosclerosis on temporal bone CT.
• Deep learning can be a helpful tool for the diagnosis of otosclerosis on temporal bone CT. • Deep learning analyses with GoogLeNet and ResNet demonstrate non-inferiority when compared to the subspecialty trained radiologist. • Deep learning may be particularly useful in medical institutions without experienced radiologists.
由于影像学表现常常不明显，因此在颞骨 CT 图像上诊断耳硬化症常常具有挑战性。我们的目的是评估深度学习分析在颞骨 CT 图像上诊断耳硬化症的效用。
共将 198 例颞骨 CT 图像分为训练集（n = 140）和测试集（n = 58）。最终诊断（耳硬化症阳性或耳硬化症阴性）由一位经验丰富的高级放射科医师确定，该医师在仔细查看所有 198 例颞骨 CT 图像的同时，结合临床和术中发现进行了诊断。在深度学习分析中，提取包含前窗裂区域的矩形目标区域，并将其输入深度学习训练过程中，以创建诊断模型。使用 AlexNet、VGGNet、GoogLeNet 和 ResNet 等深度学习模型架构进行迁移学习。随后，使用这些模型以及另一位具有 3 年神经放射学经验的放射科医师（在完成神经放射学奖学金培训后）对测试数据集进行分析。以高级放射科医师的诊断为金标准，确定放射科医师和深度学习模型的性能。
经过专门培训的放射科医师、AlexNet、VGGNet、GoogLeNet 和 ResNet 的诊断准确率分别为 0.89、0.72、0.81、0.86 和 0.86。VGGNet、GoogLeNet 和 ResNet 的性能与放射科医师无显著差异。此外，GoogLeNet 和 ResNet 与放射科医师相比表现出非劣效性。
深度学习技术可能是颞骨 CT 上诊断耳硬化症的有用辅助工具。
• 深度学习可能是颞骨 CT 上诊断耳硬化症的有用工具。• 与经过专门培训的放射科医师相比，使用 GoogLeNet 和 ResNet 的深度学习分析表现出非劣效性。• 在没有经验丰富的放射科医师的医疗机构中，深度学习可能特别有用。
