Pre-classification of Hairy Cells Using Supervised Contrastive Learning
(2022) In Master's Theses in Mathematical Sciences FMAM05 20221Mathematics (Faculty of Engineering)
- Abstract
- Differential of the peripheral blood is an important tool when assessing blood-related diseases. Lymphocytes are part of the immune system and morphological changes thereof should raise attention. CellaVision’s current systems automatically detect and pre-classify lymphocytes that exhibit atypical morphologies but do not further distinguish those into subclasses. By adding the subclasses into the pre-classification, the overall accuracy would improve and the additional information would also alleviate the work load of healthcare professionals. The objective of this thesis is to discriminate hairy cells from lymphocytes and other cell classes. The data consists of expert-labeled images of white blood cells belonging to 20 classes, and is... (More)
- Differential of the peripheral blood is an important tool when assessing blood-related diseases. Lymphocytes are part of the immune system and morphological changes thereof should raise attention. CellaVision’s current systems automatically detect and pre-classify lymphocytes that exhibit atypical morphologies but do not further distinguish those into subclasses. By adding the subclasses into the pre-classification, the overall accuracy would improve and the additional information would also alleviate the work load of healthcare professionals. The objective of this thesis is to discriminate hairy cells from lymphocytes and other cell classes. The data consists of expert-labeled images of white blood cells belonging to 20 classes, and is split into a training and a test set, with 76 612 and 15 374 images each. We compare a traditional transfer learning network against a supervised contrastive learning network and propose a methodology based on a three-step training sequence combining the two. The networks that were trained using supervised contrastive learning outperformed the traditional transfer learning networks. The best test accuracy for a contrastive learning network was 90.24%, while the best transfer learning network only obtained a test accuracy of 88.21%. With our findings, the supervised contrastive loss-aided methodology has proven to have great potential for pre-classifying hairy cells, as well as being superior in overall automatic classification of white blood cells. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9090569
- author
- Han, Alexia LU and Schömer Ericsson, Rebecca LU
- supervisor
-
- Karl Åström LU
- Kent Stråhlén LU
- Jennie Karlsson LU
- Ida Arvidsson LU
- organization
- alternative title
- Förklassning av håriga celler med hjälp av övervakad kontrastiv maskininlärning
- course
- FMAM05 20221
- year
- 2022
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- supervised contrastive learning, hairy cells, abnormal lymphocytes, deep learning, Cellavision, Xception, artificial neural network, convolutional neural networks
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMA-3473-2022
- ISSN
- 1404-6342
- other publication id
- 2022:E24
- language
- English
- id
- 9090569
- date added to LUP
- 2022-06-23 14:25:31
- date last changed
- 2022-06-23 14:25:31
@misc{9090569, abstract = {{Differential of the peripheral blood is an important tool when assessing blood-related diseases. Lymphocytes are part of the immune system and morphological changes thereof should raise attention. CellaVision’s current systems automatically detect and pre-classify lymphocytes that exhibit atypical morphologies but do not further distinguish those into subclasses. By adding the subclasses into the pre-classification, the overall accuracy would improve and the additional information would also alleviate the work load of healthcare professionals. The objective of this thesis is to discriminate hairy cells from lymphocytes and other cell classes. The data consists of expert-labeled images of white blood cells belonging to 20 classes, and is split into a training and a test set, with 76 612 and 15 374 images each. We compare a traditional transfer learning network against a supervised contrastive learning network and propose a methodology based on a three-step training sequence combining the two. The networks that were trained using supervised contrastive learning outperformed the traditional transfer learning networks. The best test accuracy for a contrastive learning network was 90.24%, while the best transfer learning network only obtained a test accuracy of 88.21%. With our findings, the supervised contrastive loss-aided methodology has proven to have great potential for pre-classifying hairy cells, as well as being superior in overall automatic classification of white blood cells.}}, author = {{Han, Alexia and Schömer Ericsson, Rebecca}}, issn = {{1404-6342}}, language = {{eng}}, note = {{Student Paper}}, series = {{Master's Theses in Mathematical Sciences}}, title = {{Pre-classification of Hairy Cells Using Supervised Contrastive Learning}}, year = {{2022}}, }