Adachi Mio, Fujioka Tomoyuki, Mori Mio, Kubota Kazunori, Kikuchi Yuka, Xiaotong Wu, Oyama Jun, Kimura Koichiro, Oda Goshi, Nakagawa Tsuyoshi, Uetake Hiroyuki, Tateishi Ukihide
Department of Surgery, Breast Surgery, Tokyo Medical and Dental University, Tokyo 113-8510, Japan.
Department of Diagnostic Radiology, Tokyo Medical and Dental University, Tokyo 113-8510, Japan.
Diagnostics (Basel). 2020 May 20;10(5):330. doi: 10.3390/diagnostics10050330.
We aimed to evaluate an artificial intelligence (AI) system that can detect and diagnose lesions of maximum intensity projection (MIP) in dynamic contrast-enhanced (DCE) breast magnetic resonance imaging (MRI). We retrospectively gathered MIPs of DCE breast MRI for training and validation data from 30 and 7 normal individuals, 49 and 20 benign cases, and 135 and 45 malignant cases, respectively. Breast lesions were indicated with a bounding box and labeled as benign or malignant by a radiologist, while the AI system was trained to detect and calculate possibilities of malignancy using RetinaNet. The AI system was analyzed using test sets of 13 normal, 20 benign, and 52 malignant cases. Four human readers also scored these test data with and without the assistance of the AI system for the possibility of a malignancy in each breast. Sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were 0.926, 0.828, and 0.925 for the AI system; 0.847, 0.841, and 0.884 for human readers without AI; and 0.889, 0.823, and 0.899 for human readers with AI using a cutoff value of 2%, respectively. The AI system showed better diagnostic performance compared to the human readers ( = 0.002), and because of the increased performance of human readers with the assistance of the AI system, the AUC of human readers was significantly higher with than without the AI system ( = 0.039). Our AI system showed a high performance ability in detecting and diagnosing lesions in MIPs of DCE breast MRI and increased the diagnostic performance of human readers.
我们旨在评估一种人工智能（AI）系统，该系统能够在动态对比增强（DCE）乳腺磁共振成像（MRI）中检测和诊断最大强度投影（MIP）的病变。我们回顾性收集了DCE乳腺MRI的MIP图像，分别用于30名正常个体和7名正常个体、49例良性病例和20例良性病例、135例恶性病例和45例恶性病例的训练和验证数据。乳腺病变用边界框表示，并由放射科医生标记为良性或恶性，而AI系统则使用RetinaNet进行训练，以检测和计算恶性可能性。使用13例正常、20例良性和52例恶性病例的测试集对AI系统进行分析。四位人类读者也对这些测试数据进行了评分，分别在有无AI系统辅助的情况下评估每个乳腺发生恶性病变的可能性。AI系统的灵敏度、特异度和受试者操作特征曲线下面积（AUC）分别为0.926、0.828和0.925；无AI辅助的人类读者分别为0.847、0.841和0.884；使用2%的临界值时，有AI辅助的人类读者分别为0.889、0.823和0.899。与人类读者相比，AI系统表现出更好的诊断性能（P = 0.002），并且由于在AI系统辅助下人类读者的性能有所提高，有AI系统时人类读者的AUC显著高于无AI系统时（P = 0.039）。我们的AI系统在检测和诊断DCE乳腺MRI的MIP病变方面表现出高性能，并提高了人类读者的诊断性能。
