Development of Tools to Classify Poor Quality Images from High Content Screening results for Cell Death and Lysosome Studies
(2020) BINP52 20192Degree Projects in Bioinformatics
- Abstract
- Cell death is an important field of biology that is linked to many other topics in medicine and biology. One of the methods to investigate cell death is through high content screening (HCS) which makes use of robotic microscopy. By this technology, cell death associated organelle membrane permeabilisation and the localisation of certain dyes can be visualised and elucidate a substantial amount of information. In HCS, an extremely large amount of images is created and much like how factories must filter out poor quality products, poor quality images must also be filtered out. Having tools for image processing and quality control available leads to easier and more accurate results in downstream processing. Here, specialised solutions were... (More)
- Cell death is an important field of biology that is linked to many other topics in medicine and biology. One of the methods to investigate cell death is through high content screening (HCS) which makes use of robotic microscopy. By this technology, cell death associated organelle membrane permeabilisation and the localisation of certain dyes can be visualised and elucidate a substantial amount of information. In HCS, an extremely large amount of images is created and much like how factories must filter out poor quality products, poor quality images must also be filtered out. Having tools for image processing and quality control available leads to easier and more accurate results in downstream processing. Here, specialised solutions were developed to detect different aberrations such as out of focus images, corrupted images, and images with dead pixels. A method for finding out of focus images was developed using the laplace operator to measure the frequency of edges in an image. In addition, machine learning was implemented using Tensorflow to train models to detect aberrant images. The best model showed an accuracy of 96% when prediction was performed on new images in the corrupted image category. An effective model could not be trained for the dead pixel category. However, this aberration type is scarce. Finally, all of the tools developed were compiled into an organised Python package for streamlined usability. (Less)
- Popular Abstract
- Filtering Good and Poor Quality Images of Cells to Research Lysosomes
With today’s technologies, we are surrounded by an immeasurable number of images that need to be analysed by computers. These images play a big part in the life sciences. Different types of images (like cells under a microscope or brain scans taken using advanced machines) can help in learning more about cells or various diseases.
The Aits Lab focuses on how cell death and a part of the cell called the lysosome can cause different diseases, so that eventually we can learn how to combat these diseases. One of the approaches that is used is analysing images of cells and lysosomes. Machine learning is an approach used to train a computer program to make predictions by... (More) - Filtering Good and Poor Quality Images of Cells to Research Lysosomes
With today’s technologies, we are surrounded by an immeasurable number of images that need to be analysed by computers. These images play a big part in the life sciences. Different types of images (like cells under a microscope or brain scans taken using advanced machines) can help in learning more about cells or various diseases.
The Aits Lab focuses on how cell death and a part of the cell called the lysosome can cause different diseases, so that eventually we can learn how to combat these diseases. One of the approaches that is used is analysing images of cells and lysosomes. Machine learning is an approach used to train a computer program to make predictions by having it learn from data. At present it is a very popular way to find patterns in science. An example of how machine learning can be used is to have a computer program that can tell you if a picture has a dog or a cat. In our study, we want to know if the image is a good quality image or a poor quality image. An example of the images we work with is shown below, with different parts of the cell represented by different colours.
Consider a person learning how to make fried rice. In order to do this, the person reads 100 fried rice recipes. However, imagine that in 5 of the recipes, there was a typo where it said “mice” instead of “rice.” This might confuse the person but since there are only five instances, they would assume it’s a typo. However, if 60 of the recipes said “mice”, then they might end up making fried mice instead of fried rice. In the same way, if we’re teaching machines to do something and the data is of poor quality, the machine won’t learn the way we want it to.
Testing Methods for Three Different Error Types
Different approaches were evaluated for three different error types: blurry images, corrupted images, and images with dead pixels. An accurate method was developed for the blurry images involving measuring how blurry an image was and then setting a threshold which would exclude images that were too blurry. Although it can still be improved, it was able to predict blurry images almost perfectly with the right settings. For corrupted images, a machine learning model was trained to predict corrupted images and had over 95% accuracy. We experimented with different methods for detecting the dead pixel error, but none gave good results.
These methods can be further developed and then used to find and remove poor quality images from our data set. This way, when we’re working with images later on, we’ll know we’re working with good quality images.
Master’s Degree Project in Bioinformatics BINP52 credits 60
Department of Biology, Lund University
Advisor: Dr. Sonja Aits
Advisors Department of Experimental Medicine (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9031410
- author
- Chan Lim, Jong
- supervisor
-
- Sonja Aits LU
- organization
- course
- BINP52 20192
- year
- 2020
- type
- H2 - Master's Degree (Two Years)
- subject
- language
- English
- id
- 9031410
- date added to LUP
- 2020-10-27 16:06:23
- date last changed
- 2020-10-27 16:06:23
@misc{9031410,
abstract = {{Cell death is an important field of biology that is linked to many other topics in medicine and biology. One of the methods to investigate cell death is through high content screening (HCS) which makes use of robotic microscopy. By this technology, cell death associated organelle membrane permeabilisation and the localisation of certain dyes can be visualised and elucidate a substantial amount of information. In HCS, an extremely large amount of images is created and much like how factories must filter out poor quality products, poor quality images must also be filtered out. Having tools for image processing and quality control available leads to easier and more accurate results in downstream processing. Here, specialised solutions were developed to detect different aberrations such as out of focus images, corrupted images, and images with dead pixels. A method for finding out of focus images was developed using the laplace operator to measure the frequency of edges in an image. In addition, machine learning was implemented using Tensorflow to train models to detect aberrant images. The best model showed an accuracy of 96% when prediction was performed on new images in the corrupted image category. An effective model could not be trained for the dead pixel category. However, this aberration type is scarce. Finally, all of the tools developed were compiled into an organised Python package for streamlined usability.}},
author = {{Chan Lim, Jong}},
language = {{eng}},
note = {{Student Paper}},
title = {{Development of Tools to Classify Poor Quality Images from High Content Screening results for Cell Death and Lysosome Studies}},
year = {{2020}},
}