The Berkeley Single Cell Computational Microscopy (BSCCM) Dataset

Henry Pinkard, Cherry Liu, Fanice Nyatigo, Daniel A. Fletcher, and Laura Waller

Abstract

Computational microscopy, in which hardware and algorithms of an imaging system are jointly designed, shows promise for making imaging systems that cost less, perform more robustly, and collect new types of information. Often, the performance of computational imaging systems, especially those that incorporate machine learning, is sample-dependent. Thus, standardized datasets are an essential tool for comparing the performance of different approaches. Here, we introduce the Berkeley Single Cell Computational Microscopy (BSCCM) dataset, which contains ~12,000,000 images of 400,000 individual white blood cells. The dataset contains images captured with multiple illumination patterns on an LED array microscope and fluorescent measurements of the abundance of surface proteins that mark different cell types. We hope this dataset will provide a valuable resource for the development and testing of new algorithms in computational microscopy and computer vision with practical biomedical applications.

Resources

Overview

Figure 1
Berkeley Single Cell Computational Microscopy (BSCCM) dataset overview. a) Schematic of the microscope used in data collection: a com- mercial body fluorescence microscope with its trans-illumination lamp replaced with a programmable LED array quasi-dome. b) The LED array was used for label-free imaging of cells with different illumination patterns. c) The fluorescence light path was used capture 6-channel fluorescence images on the same cells. d) This provided both protein expression levels and label-free images on the same cells.
Figure 1
LED array and histology contrasts in BSCCM and BSCCM- coherent

BSCCM variants

Name Size (GB) # Cells Image Specifications
LED-array/Fluor
Histology
BSCCM 228 412,941 128×128 12-bit
398×398 36-bit
BSCCMNIST 58 412,941 28×28 8-bit
28×28 24-bit
BSCCM-coherent 30 4,304 128×128 12-bit
BSCCM-tiny 0.6 1,000 128×128 12-bit
398×398 36-bit
BSCCMNIST-tiny 0.2 1,000 28×28 8-bit
28×28 24-bit
BSCCM-coherent-tiny 0.6 100 128×128 12-bit