Multiplane microscopy dataset for benchmarking denoising methods

The dataset consists of 810 microscopy images collected from CHO, U2OS, and RPE1 cell lines in fluorescent and brightfield modalities. Cell lines were grown in standard medium, seeded on plates, fixed with formaldehyde, stained with a fluorescent dye (Hoechst33342) and imaged with PerkinElmer Phenix high-throughput confocal microscope in fluorescence and brightfield modalities using a 20x objective. For the fluorescent modality, images were acquired in low-exposure (20 ms) and high-exposure (100 ms) modes. For the brightfield, we used a single exposure (100 ms). Higher exposure time usually translates into better image quality, alleviating Poisson noise, however, it leads to sample degradation, and lower measurement speed. In total, three wells on a plate were imaged, one well per cell line, each with nine fields of view. Each field of view was imaged across ten focal planes separated by 2 micrometers in z-stack, resulting in 270 microscopy images of size 2160 x 2160 pixels per exposure time per modality. We used the following folder structure: /{well}_{cell line}/{modality}_{exposure time}/fov_{field of view}_plane_{plane}.png We recommend to use the first two wells (CHO and U2OS, 180 images) as a training set, the first three fields of view from the last well (RPE1, 30 images) as a validation set, and the last six fields of view (RPE1, 60 images) as a test set.本数据集共包含810幅显微图像，采集自CHO、U2OS与RPE1三种细胞系，涵盖荧光成像模式（fluorescent modality）与明场成像模式（brightfield modality）。所有细胞系均采用标准培养基培养，接种于培养板后经甲醛固定、荧光染料Hoechst33342染色，随后使用PerkinElmer Phenix高通量共聚焦显微镜，以20倍物镜分别采集两种成像模式下的图像。 荧光成像模式下，图像分别以低曝光时长（20 ms）与高曝光时长（100 ms）两种模式采集；明场成像模式下则采用单次曝光（100 ms）。曝光时长越长，图像质量通常越好，可有效缓解泊松噪声（Poisson noise），但同时会导致样本降解并降低成像速度。 实验中共对一块培养板的3个孔进行成像，每个细胞系对应1个孔，每个孔包含9个视场（field of view）。每个视场均通过Z轴堆叠（z-stack）采集10个焦距平面，相邻平面间距为2微米；针对每种曝光时长与成像模式，均可得到270幅分辨率为2160×2160像素的显微图像。 本数据集采用如下文件夹组织结构：/{well}_{cell line}/{modality}_{exposure time}/fov_{field of view}_plane_{plane}.png 我们建议将前两个孔（对应CHO与U2OS细胞系，共180幅图像）作为训练集（training set）；将最后一个孔（对应RPE1细胞系）的前3个视场（共30幅图像）作为验证集（validation set）；将该孔的后6个视场（共60幅图像）作为测试集（test set）。