Machine Learning Features from Proton Therapy Treatment Simulations with the Bergen DTC Prototype for Range Verification

Extracted features from the simulation data found at DOI: 10.5281/zenodo.8192778 Each simulation constitutes a single data sample. The following features were extracted. Detector features: Total number of active pixels Total number of clusters (hits) Number of clusters over threshold (5, 20 pixels) Mean and standard deviation of cluster sizes The number of clusters of any given size (1–72) Mean and standard deviation of x- and y-coordinates over each layer (0–42), and the entire detector Number of active pixels in each layer (0–42) Number of clusters (hits) in each layer (0–42) Total energy deposition of the hits in each layer (0–42) Higher-level detector features, i.e., function fits (linear, cubic, exponential) with their mean squared residuals over the following quantities: Active pixels over layer Number of clusters over layer Total deposited energy over layer 201 RSP features extracted from the beam spot, the phantom rotation, and its 3D RSP image. Two datasets are included in two separate archive files: <strong>features.tar.gz:</strong> 715-HN phantom by CIRS Inc. (Norfolk, VA, United States), digitized by Giacometti et al. (2017). <strong>features-vhf.tar.gz:</strong> The Visible Human Female (VHF) Head phantom (Ackermann et al. 1995), courtesy of the U.S. National Library of Medicine, resampled to 1 mm voxels and scaled down to 80% size in the simulation. After extracting features, some outliers were removed from the datasets: 14 samples for 715-HN and 3 samples for VHF. The rest of the samples were split into train (70%), validation (10%), and test (20%) sets, for both phantoms separately, which can be found in separate CSV files: features_train.csv, features_val.csv, features_test.csv (715-HN) and features-vhf_train.csv, features-vhf_val.csv, features-vhf_test.csv (VHF). The last file (features_shifted_test.csv (715-HN) and features-vhf_shifted_test.csv (VHF)) contains 40 additional samples for each data point in the respective test set, representing a simulated lateral shift between 1 mm and 10 mm in 1 mm intervals in all directions along the x- and y-axis of the beam.本数据集的特征提取自DOI: 10.5281/zenodo.8192778 公开的模拟数据，每份模拟结果对应一个独立数据样本。本次提取的特征如下： ### 探测器特征 1. 有效像素（active pixels）总数 2. 簇（clusters，即击中事件）总数 3. 超过阈值（5、20像素）的簇数量 4. 簇尺寸的均值与标准差 5. 任意给定尺寸（1~72）的簇数量 6. 各层（0~42层）以及整个探测器上x、y坐标的均值与标准差 7. 各层（0~42层）的有效像素数量 8. 各层（0~42层）的簇（击中事件）数量 9. 各层（0~42层）的击中总能量沉积 ### 高阶探测器特征 即针对以下物理量进行的函数拟合（线性、三次、指数拟合）及其均方残差（mean squared residuals）： - 各层有效像素数 - 各层簇数量 - 各层总沉积能量 ### 束流相关RSP特征 从束斑（beam spot）、体模（phantom）旋转及其三维相对停止功率（Relative Stopping Power, RSP）图像中提取的201项RSP特征。 本数据集包含两个独立归档文件： - <strong>features.tar.gz:</strong> 由美国弗吉尼亚州诺福克市CIRS公司生产的715-HN体模，经贾科梅蒂等人（2017）数字化处理。 - <strong>features-vhf.tar.gz:</strong> 可视女性人体（Visible Human Female, VHF）头部体模（阿克曼等人，1995），由美国国家医学图书馆提供，已重采样至1mm体素并在模拟中按80%比例缩放。 完成特征提取后，我们对数据集进行了异常值剔除：715-HN体模数据集移除14个样本，VHF头部体模数据集移除3个样本。针对两个体模的剩余样本，分别按70%训练集、10%验证集、20%测试集的比例划分为三个子集，对应存储于独立的CSV文件中： - 715-HN体模：features_train.csv、features_val.csv、features_test.csv - VHF头部体模：features-vhf_train.csv、features-vhf_val.csv、features-vhf_test.csv 另有两组额外测试集文件（715-HN体模对应features_shifted_test.csv，VHF头部体模对应features-vhf_shifted_test.csv），每个原测试集的数据点均对应40个额外样本，代表沿束流x轴与y轴各方向以1mm为间隔，在1mm至10mm范围内的模拟横向偏移（lateral shift）场景。