University of Patras, Medical School, Department of Medical Physics, Rio, Achaia, PC 26504, Greece.
Hell J Nucl Med. 2020 May-Aug;23(2):125-132. doi: 10.1967/s002449912101. Epub 2020 Jul 27.
To investigate a deep learning technique, more specifically state-of-the-art convolutional neural networks (CNN), for automatic characterization of polar maps derived from myocardial perfusion imaging (MPI) studies for the diagnosis of coronary artery disease.
Stress and rest polar maps corresponding to 216 patient cases from the database of the department of Nuclear Medicine of our institution were analyzed. Both attenuation-corrected (AC) and non-corrected (NAC) images were included. All patients were subjected to invasive coronary angiography within 60 days from MPI. As the initial dataset of this study was small to train a deep learning model from scratch, two strategies were followed. The first is called transfer learning. For this, we employed the state-of-the-art CNN called VGG16, which has been broadly exploited in medical imaging classification tasks. The second strategy involves data augmentation, which is achieved by the rotation of the polar maps, to expand the training set. We evaluated VGG16 with 10-fold cross-validation on the original set of images performing separate experiments for AC and NAC polar maps, as well as for their combination. The results were compared to the standard semi-quantitative polar map analysis based on summed stress and summed difference scores, as well as to the medical experts' diagnostic yield.
With reference to the findings of coronary angiography, VGG16 achieved an accuracy of 74.53%, sensitivity 75.00% and specificity 73.43% when the AC and NAC polar maps were incorporated into one single image set. Respective figures of MPI interpretation by experienced Nuclear Medicine physicians were 75.00%, 76.97% and 70.31%. The accuracy of semi-quantitative polar map analysis was lower, 66.20% and 64.81% for AC and NAC technique, respectively.
The proposed deep learning model with data augmentation techniques performed better than the conventional semi-quantitative polar map analysis and competed with doctor's expertise in this particular patient cohort and image set. The model could potentially serve as an assisting tool to support interpretation of MPI studies or could be used for teaching purposes.
研究一种深度学习技术，特别是最先进的卷积神经网络（CNN），用于自动对心肌灌注成像（MPI）研究得出的极坐标图进行特征描述，以诊断冠状动脉疾病。
分析了来自我们机构核医学科数据库的 216 例患者的应激和静息极坐标图。包括校正衰减（AC）和未校正（NAC）图像。所有患者均在 MPI 后 60 天内行有创冠状动脉造影。由于本研究的初始数据集太小，无法从头开始训练深度学习模型，因此采用了两种策略。第一种是迁移学习。为此，我们使用了一种名为 VGG16 的最先进的 CNN，该 CNN 已广泛应用于医学影像分类任务中。第二种策略涉及数据扩充，通过极坐标图的旋转来扩大训练集。我们在原始图像集上进行了 10 折交叉验证，对 AC 和 NAC 极坐标图分别进行了实验，并对它们的组合进行了实验。结果与基于总和应激和总和差异评分的标准半定量极坐标图分析以及医学专家的诊断率进行了比较。
参考冠状动脉造影的结果，当将 AC 和 NAC 极坐标图合并到一个单一的图像集中时，VGG16 的准确率为 74.53%，灵敏度为 75.00%，特异性为 73.43%。有经验的核医学医师对 MPI 的解释分别为 75.00%、76.97%和 70.31%。半定量极坐标图分析的准确率较低，AC 和 NAC 技术的准确率分别为 66.20%和 64.81%。
在该特定患者队列和图像集中，与传统的半定量极坐标图分析相比，该方法具有数据扩充技术的深度学习模型表现更好，并与医生的专业知识相竞争。该模型可以作为辅助解释 MPI 研究的工具，也可以用于教学目的。
