Domain adaptation for white blood cell classification - Bridging the gap created by optical hardware upgrades
(2022) In Master’s Theses in Mathematical Sciences FMAM05 20221Mathematics (Faculty of Engineering)
- Abstract
- Collecting enough labeled images to train an artificial neural network (ANN) is a common struggle since the process is tedious, time consuming and expensive. CellaVision, a company that creates analyzers that automate the process of finding and pre-classifying abnormal blood cells using machine learning, runs into this problem regularly. When new generations of analyzers are being produced, the images they capture will usually have different hues compared to the previous generation due to the optical hardware upgrades. This means that the ANNs used to pre-classify the cell types need to be re-trained with images captured by the new analyzer.
One possible way of reducing the amount of new data that needs to be collected and labeled is by... (More) - Collecting enough labeled images to train an artificial neural network (ANN) is a common struggle since the process is tedious, time consuming and expensive. CellaVision, a company that creates analyzers that automate the process of finding and pre-classifying abnormal blood cells using machine learning, runs into this problem regularly. When new generations of analyzers are being produced, the images they capture will usually have different hues compared to the previous generation due to the optical hardware upgrades. This means that the ANNs used to pre-classify the cell types need to be re-trained with images captured by the new analyzer.
One possible way of reducing the amount of new data that needs to be collected and labeled is by using domain adaptation. Domain adaptation is, in this case, using the already existing and labeled images from the older generation of analyzers to improve the performance of the networks built to pre-classify the images of cells from the new analyzers.
This thesis seeks to answer whether images from older generations of analyzers can be used to train an ANN to pre-classify images of white blood cells captured by a new analyzer. Various machine learning methods have been tried to find the most effective way to apply domain adaptation in this setting. After a robust evaluation using K-fold cross-validation, we found that the methods we used for domain adaptation may only give marginal improvement when applied in this context. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9086197
- author
- Lindhé, Elsa LU and Olsson, Erik
- supervisor
- organization
- course
- FMAM05 20221
- year
- 2022
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Domain adaptation, leukocytes, WBC, machine learning, deep learning
- publication/series
- Master’s Theses in Mathematical Sciences
- report number
- LUTFMA-3468-2022
- ISSN
- 1404-6342
- other publication id
- 2022:E18
- language
- English
- id
- 9086197
- date added to LUP
- 2022-06-17 16:57:46
- date last changed
- 2022-06-17 16:57:46
@misc{9086197, abstract = {{Collecting enough labeled images to train an artificial neural network (ANN) is a common struggle since the process is tedious, time consuming and expensive. CellaVision, a company that creates analyzers that automate the process of finding and pre-classifying abnormal blood cells using machine learning, runs into this problem regularly. When new generations of analyzers are being produced, the images they capture will usually have different hues compared to the previous generation due to the optical hardware upgrades. This means that the ANNs used to pre-classify the cell types need to be re-trained with images captured by the new analyzer. One possible way of reducing the amount of new data that needs to be collected and labeled is by using domain adaptation. Domain adaptation is, in this case, using the already existing and labeled images from the older generation of analyzers to improve the performance of the networks built to pre-classify the images of cells from the new analyzers. This thesis seeks to answer whether images from older generations of analyzers can be used to train an ANN to pre-classify images of white blood cells captured by a new analyzer. Various machine learning methods have been tried to find the most effective way to apply domain adaptation in this setting. After a robust evaluation using K-fold cross-validation, we found that the methods we used for domain adaptation may only give marginal improvement when applied in this context.}}, author = {{Lindhé, Elsa and Olsson, Erik}}, issn = {{1404-6342}}, language = {{eng}}, note = {{Student Paper}}, series = {{Master’s Theses in Mathematical Sciences}}, title = {{Domain adaptation for white blood cell classification - Bridging the gap created by optical hardware upgrades}}, year = {{2022}}, }