Image dataset for a CNN algorithm development to detect coronary atherosclerosis in coronary CT angiography

Coronary artery image sets for 500 patients. Each image represents a Mosaic Projection View (MPV) which consists of 18 different views of a straightened coronary artery stacked vertically. Training-Validation-Test image sets are sub-divided per patient with 3/1/1 ratio (300/100/100), each with 50% normal and 50% diseased cases. Artery images derived from the 300 training cases were augmented 6-fold to create 2,364 (i.e., 394 x 6) images in order to strengthen modeling and dataset balance. However, augmentation was not performed on the: 1. normal component of the training dataset (2,304 images); 2. entire validation dataset; or 3. entire testing dataset. In the validation dataset, only one artery was randomly selected per normal case (50 images) and diseased case (50 images) for balance maintenance.

本冠状动脉影像数据集包含500名患者的样本。每张影像均为镶嵌投影视图（Mosaic Projection View, MPV），该视图由垂直堆叠的18幅经拉直处理的冠状动脉不同视角影像拼接而成。训练集、验证集与测试集按照患者个体进行划分，划分比例为3:1:1（分别对应300、100、100名患者），且每个子集内正常病例与病变病例占比均为50%。从300例训练病例中提取的动脉影像经6倍数据增强后，共生成2364幅影像（即394×6），以强化模型训练效果并平衡数据集。但以下三类对象未实施数据增强操作：1. 训练数据集的正常病例子集（共2304幅影像）；2. 完整的验证数据集；3. 完整的测试数据集。在验证数据集中，为维持数据平衡，仅从每例正常病例（对应50幅影像）与每例病变病例（对应50幅影像）中随机选取一条动脉的影像。