Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

3D Image Analysis of Oat Seeds using Deep Learning

Roostee, Suze Julia (2020) BINP51 20192
Degree Projects in Bioinformatics
Popular Abstract
Understanding Oat with Artificial Intelligence

Eating or drinking oat is considered healthy for you and the planet, but there is still much scientists don’t know about oat. One of the things they would really like to know is how exactly oat stores its molecules and how the structure of oat is related to that. One way to do this is by scanning oats with CT (a bit like a brain scan in the hospital) to image them while keeping their structure intact. By stacking hundreds to thousand of 2D images that come out of the CT scan a 3D view of the oat seed arises, and the scientists can start to learn about the structure. An important step in this process is to assign what part of the image relates to what part of the internal oat structure, such... (More)
Understanding Oat with Artificial Intelligence

Eating or drinking oat is considered healthy for you and the planet, but there is still much scientists don’t know about oat. One of the things they would really like to know is how exactly oat stores its molecules and how the structure of oat is related to that. One way to do this is by scanning oats with CT (a bit like a brain scan in the hospital) to image them while keeping their structure intact. By stacking hundreds to thousand of 2D images that come out of the CT scan a 3D view of the oat seed arises, and the scientists can start to learn about the structure. An important step in this process is to assign what part of the image relates to what part of the internal oat structure, such as the germ, endosperm, and aleurone (see figure). This requires manual labelling of every pixel in every image, a laborious and time-consuming task. Can’t the computer learn to do this?


Nowadays Artificial Intelligence (AI) has applications in many fields. From face recognition in social media to object detection in cars, AI surrounds us. With the faster computers and large graphics cards that are available these days image recognition is one of the fields that is doing particularly well. And within the image recognition field a specific type of models, called Convolutional Neural Networks (CNNs) are especially successful.

CNNs are made up of layers of calculations, where the output of a layer is the input to a next layer. When we start to teach the model something about the images, like what class every pixel belongs to, we provide these images as input to the model. However, the CNN doesn’t see an image like we do, instead it sees rows and columns filled with numbers, where every pixel in the image is a number. The CNN wants to give the right answer to the question ‘What class is this pixel?’ and tries to create a mathematical function that can calculate the right answer. Generally the model needs to see thousands of images and their correct answer to learn a function that is able to give the correct answer on images the CNN hasn’t seen before.

Diving a bit deeper into the structure of the CNN and the layers it has, shows us that different layers in the model can recognise different structures in the images. The first layers in the model can learn to recognise easy structures, such as lines, borders, and circles. When the model has more layers the deeper layers can combine the knowledge of the first layers in ‘complex’ structures, like faces. Knowing this gives us a nice advantage that can speed up training of the CNN, and even make the predictions by the model more accurate. Instead of training a CNN that has to recognise oat seed images from scratch we can copy some of the first layers from another CNN. This CNN has been trained on ImageNet, a database with 14 million images, and is very good at recognising structures. When we take the first layers of this CNN and put it into ours, we only have to teach our CNN how to recognise the oat seed structures from that.

And, can it?

It turns out that CNN's can also learn to recognise different structures in an oat seed, and that this can be used to speed up oat seed image analysis. Using this CNN may allow for easier and faster screening of hundreds of oat seed images. That way scientists will be able to look into genetic varieties of oat seeds, and possibly implement this for oat seed selection in oat breeding programs.

Master’s Degree Project in Bioinformatics 45 credits 2019
Department of Biology, Lund University

Advisors: Nick Sirijovski & Nikos Tsardakas Renhuldt
Pure & Applied Biochemistry, Lunds Tekniska Högskola (Less)
Please use this url to cite or link to this publication:
author
Roostee, Suze Julia
supervisor
organization
course
BINP51 20192
year
type
H2 - Master's Degree (Two Years)
subject
language
English
id
9022014
date added to LUP
2020-06-25 12:29:46
date last changed
2020-06-25 12:29:46
@misc{9022014,
  author       = {{Roostee, Suze Julia}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{3D Image Analysis of Oat Seeds using Deep Learning}},
  year         = {{2020}},
}