Dataset: "Deep learning with digital holographic microscopy discriminates apoptosis and necroptosis"

In this repository you will find all data, models and MATLAB code used with this publication; DOI: (pending). This is used for cell death detection on DHM (Digital holographic microscopy) data using deep learning in MATLAB. This processing pipeline consists of a few different parts. 1) SAD.m Supervised Anomaly Detection in order to remove all Alive cells from cell death experiments and vice versa. This since cell death induction (like any other biological process) is happening not at the same time; the cell population is heterogeneous. SAD filters and homogenizes the dataset, the SAD filter is stored in Models/SAD_filter.mat 2) CropFromCaptureAndFilter.m This script takes the captures (whole field of view, from folder Captures) of the DHM camera snaps and crops out the cells from each capture. The single cells are stored in CroppedImages 3) These CroppedImages are split in HoldoutSet and TrainingSet 4) MakeModelViaTransferLearning.m Uses Transfer learning to reuse the VGG-19 network and repurpose this to predict cell death on these data. This script relearns the model on the trainingset from the folder TrainingSet and saves the model under Models/convnet.mat 5) UseModelOnHoldoutSet.m Opens ConvNet model (created by MakeModelViaTransferLearning.m) and the HoldoutSet (independent experiments) and runs the model on these images. Later on the overall accuracy is measured (total correct predictions/total samples *100%) and a confusion matrix is made. Finally an ROC plot is made.

本仓库包含本出版物所使用的全部数据、模型与MATLAB代码；数字对象唯一标识符（DOI）尚在审核中。本项目基于MATLAB平台的深度学习方法，用于对数字全息显微镜（Digital Holographic Microscopy, DHM）采集的数据开展细胞死亡检测任务。整套处理流程包含以下五个环节：1) SAD.m：即监督异常检测（Supervised Anomaly Detection, SAD）脚本，用于在细胞死亡实验中剔除存活细胞，反之亦可。由于细胞死亡诱导过程与其他生物过程一致，并非同步发生，细胞群体具有异质性，因此SAD模块可对数据集进行滤波与同质化处理，该SAD滤波器存储于Models/SAD_filter.mat路径下。2) CropFromCaptureAndFilter.m脚本：该脚本从Captures文件夹中读取DHM相机拍摄的全视场图像，从每幅图像中裁剪出单个细胞，并将裁剪得到的单个细胞图像存储至CroppedImages文件夹。3) 将上述CroppedImages文件夹中的细胞图像划分为留出集（HoldoutSet）与训练集（TrainingSet）。4) MakeModelViaTransferLearning.m脚本：通过迁移学习（Transfer Learning）复用VGG-19网络，将其适配至本数据集的细胞死亡预测任务。该脚本在TrainingSet文件夹的训练数据上重新训练模型，并将训练完成的模型保存至Models/convnet.mat路径。5) UseModelOnHoldoutSet.m脚本：加载由MakeModelViaTransferLearning.m生成的ConvNet模型与留出集（独立实验数据集），并在这批图像上运行模型。随后计算整体准确率（总正确预测数/总样本数×100%），生成混淆矩阵，最终绘制受试者工作特征（Receiver Operating Characteristic, ROC）曲线。