Proton2Carbon dataset: pairs of proton and carbon ion pencil beams images obtained using a scintillation detector

The Proton2Carbon dataset contains 547,224 pairs of proton and carbon ion pencil beams images obtained using a scintillation detector. Each image represents a pulled back pencil beam that crossed an object located between the source and the detector at a given location in the imaging field of view. All details related to image acquisition can be found at https://doi.org/10.1002/mp.17645. Each of the 547,224 images is a (W, H) = (608, 96) pixels .png image containing a (pulled back) pencil beam, with various levels of range mixing. The camera pixel size is approximately 0.41 mm; each image covers a physical field of view of approximately 250x40 mm2. While the scintillator acquires raw images of size (608, 488), each image was cropped to a 4 cm region around the centre of the pencil beam in the height dimension to reduce the size of the images and network. For users interested in image reconstruction, we provide the coordinate of the crop for each image in the cropped_coordinates.csv file such that the full (608, 488) images can be recreated. To recreate the full image, assuming that cropped_image is the (608, 96) image, one can do full_image = np.zeros((608, 488)) full_image[yi:yf, xi:xf] = cropped_image Where xi, xf, yi, yf are tabulated in the .csv file. The dataset contains 12 zipped folders, one for each of the 12 scans introduced in section 1. The data structure is: ./<scan_number>/<ion_species>/<camera_view>_image_#.png Where <scan_number> {scan_1, …, scan_12} and correspond to the 12 geometries introduced in section 1, <ion_species> {protons, carbon}, <camera_view> {lateral, top}, and # represents the number of the image, ranging from 1 to 22,801 for each combination of scan, ion species and camera view. The matching proton and carbon ion images will have the same filename except the <ion_species>.Proton2Carbon数据集（Proton2Carbon Dataset）包含547,224对质子与碳离子笔形束图像，此类图像通过闪烁探测器（scintillation detector）采集得到。每幅图像对应一束经拉远处理的笔形束，该笔形束在成像视场的指定位置穿过源与探测器之间的物体。所有与图像采集相关的详细信息均可查阅https://doi.org/10.1002/mp.17645。 547,224幅图像均为尺寸(W, H) = (608, 96)像素的PNG图像，内容为经拉远处理的笔形束，且存在不同程度的射程混合。相机像素尺寸约为0.41 mm，单幅图像覆盖的物理视场约为250×40 mm²。闪烁探测器采集的原始图像尺寸为(608, 488)，为压缩图像体积与神经网络模型规模，我们将图像在高度维度上裁剪至笔形束中心周围4 cm的区域，最终得到(608, 96)尺寸的图像。 针对有图像重建需求的用户，我们在cropped_coordinates.csv文件中提供了每幅图像的裁剪坐标，以便还原完整的(608, 488)尺寸原始图像。若要还原完整图像，假设cropped_image为(608, 96)的裁剪后图像，可通过如下代码实现： full_image = np.zeros((608, 488)) full_image[yi:yf, xi:xf] = cropped_image 其中xi、xf、yi、yf的取值均已在上述CSV文件中列出。 本数据集包含12个压缩文件夹，分别对应第1节中介绍的12次扫描。数据集的目录结构如下： ./<scan_number>/<ion_species>/<camera_view>_image_#.png 其中<scan_number>取值为{scan_1, …, scan_12}，对应第1节中介绍的12种扫描几何结构；<ion_species>取值为{protons, carbon}，即质子与碳离子；<camera_view>取值为{lateral, top}，即侧视与顶视；#代表图像编号，每种扫描、离子种类与相机视角的组合对应1至22,801号图像。匹配的质子与碳离子图像仅在<ion_species>字段存在差异，其余文件名完全一致。