Trypan Blue stained Cells Image Dataset

This Image Dataset contains 39
real world color images of trypan blue stained animal cells. Out of the 39
images, 21 were acquired using a digital camera mounted on an EclipseE200
(Nikon, Japan) bright field optical microscope with a 4x objective lens. For
these images, NIH 3T3 cells were cultured, mixed on a 1:1 ratio with 0.4%
trypan blue and loaded onto a Neubauer Chamber (Marienfeld, Germany). The field
of view of each image is of approximately 2 mm × 1.5 mm, and its resolution of
2592×1936 pixels. The rest of the images were obtained from Chan, L. L.-Y., Rice, W. L. & Qiu, J.
(2020). Observation and quantification of the morphological effect of trypan
blue rupturing dead or dying cells. Plos one, 15(1), e0227950.

Real world images were used to
create a larger Synthetic Image Dataset. In order to do so, single-cell masks
were obtained from each image and classified according to its state: live or
dead. Also, from each image, masks of cell clusters and debris were obtained
and separated from its background.

The synthetic image generation
process is described as follows. First, a random background was selected from the
background image pool and was resized to occupy a total of 3280x2464 pixels. Second,
live and dead cell masks were randomly selected from the live and dead image
pools and were pasted on top of the background using random x and y coordinates. Backgrounds, as well as live and dead masks, were randomly
flipped (vertically) and/or mirrored (horizontally) and modified in its
brightness, contrast and sharpness when generating each image. Cell masks were
also rotated. Bounding box annotations were included in individual .txt files
using the YOLO format and were automatically generated simultaneously with the
image synthesis. A total of 2192 training images and 250 validation images were
generated.

Single-cell masks and backgrounds
used for the image synthesis were obtained from 24 of the 39 real world images.
The remaining 15 images (8, 16, 19, 21-25 & 33-39) were used for model testing.

The trained YOLOv4 cell counting
model obtained a mAP50 of 87.30%, 88.47% of Precision and 90.24% of
Recall in real world images. This model was used for the development of a
stand-alone, portable and low-cost Automated cell-counter.