Medical Image Analysis -
Automated Embryo Mitosis Detection


During IVF (In-Vitro Fertilization), human oocytes (unfertilized eggs) are fertilized In-Vitro (in glass, outside body). The resulting embryos (fertilized eggs) are grown in an incubator for 3-5 days before they are re-implanted into the patient’s uterus. During the first 3 days, starting from single cell embryos, the cells divide approximately 3 times into 8-cell embryos (1-2-4-8). The timings of these cell divisions (mitosis) have been determined to be important indicators for pregnancy outcomes.

The incubators, where embryos are grown, have built-in microscopes with cameras taking pictures every 15 minutes. These pictures can be combined into fast-forwarded movie sequences (time-lapse photography) from which cell divisions can be observed. The objective is to analyze these images, count the number of cells in each image, to determine exactly when mitosis occured.

Cicconet et al. [1] have developed an automated method to detect mitosis up to 4-cell stage, and semi-automated up to 8-cell stage, in mouse embryos. They have also published a database with time-lapse image sequences of 100 mouse embryos. Mouse embryos develop similar to human embryos, but faster, and more consistent with regards to shape, making them easier to analyze.


For this project, we have access to a larger dataset of mouse embryo images. The objective is to repeat Cicconet’s research using our dataset, and if possible improve the method. Cicconet has made their source code available at It's written using Objective-C, and will only run on a Mac. It would be nice to have it translated into a language that can run on a PC with Windows and/or Linux, for example Java, Python, or MATLAB.

Extended Project

Given that we have a large dataset, it may be suitable for a deep learning approach. Either directly using unsupervised learning, or first using some method of labeling the images, and then using a supervised learning method.


Embryos were incubated using Embryoscope machines from Vitrolife. They work by inserting slides of 12 embryos each, which are then photographed every 15 minutes over 32.5 hours, resulting in 130 images of each embryo. We have a datased consisting of 234 slides, that is 2808 embryos, or 365040 images. Each image is also taken 3 times with different camera focus providing some depth. Including these we have over 1 million images, or about 25.8 Gb of data. Each image, see examples below, is 500x500 pixels and compressed with jpeg. One of the first steps in this project should be to crop all images to eliminate parts that are not of embryos, and reduce the datasize whithout losing relevant information.

Single cell embryo at start of incubation

2-cell embryo after 19 15-minute intervals (4h 45m)

3-cell embryo after 20h 45m

4-cell embryo after 21h 15m


  1. Cicconet, M., et al. (2014). "Label Free Cell-Tracking and Division Detection Based on 2D Time-Lapse Images For Lineage Analysis of Early Embryo Development." Computers in biology and medicine 51: 24-34.