Deep Learning for Breast Cancer Detection in Ultrasound Imaging : Classification, Managing Data Scarcity and Detecting Out-of-Distribution Samples
(2024) In Licentiate theses in mathematical sciences 2024(2).- Abstract
- Breast cancer has a profound affect on society. The survival for women in low- and middle-income countries (LMICs) is poor compared to in high-income countries (HICs). The lack of timely diagnosis is one of the main factors contributing to the poor outcomes for women in LMICs. Point-of-care ultrasound (POCUS) combined with a deep learning (DL) classification network could potentially be a suitable support tool for breast cancer detection in LMICs.
There are different ways of designing a classification network. Convolutional neural networks (CNNs) are widely used for image classification tasks and are becoming regular in medical applications. To train a DL classification network data is needed. Medical data can be difficult to... (More) - Breast cancer has a profound affect on society. The survival for women in low- and middle-income countries (LMICs) is poor compared to in high-income countries (HICs). The lack of timely diagnosis is one of the main factors contributing to the poor outcomes for women in LMICs. Point-of-care ultrasound (POCUS) combined with a deep learning (DL) classification network could potentially be a suitable support tool for breast cancer detection in LMICs.
There are different ways of designing a classification network. Convolutional neural networks (CNNs) are widely used for image classification tasks and are becoming regular in medical applications. To train a DL classification network data is needed. Medical data can be difficult to acquire due to many different reasons, one of them being ethical approvals. The availability of breast POCUS data is limited. However, the access to standard ultrasound (US) images is greater. There are different methods to expand a data set, a very common one is the use of data augmentation. Another interesting method is the cycle-consistent adversarial network (CycleGAN). This network is trained to transform an image of one domain into another domain. Thus, standard US images could be transformed into the domain of POCUS imaging, generating more POCUS samples without the need of collecting the POCUS images in the clinic. Further, it is crucial to assure the classification network to be trustworthy. Images which are out-of-distribution (OOD) should be detected and no prediction by the network should be made. OOD samples in breast US imaging includes; images of poor quality, images capturing other structures than breast tissue and images showing rare lesions. For the third case, rare lesions, the sample is in-distribution (ID). However, since the lesion is rare and might not be covered within the knowledge of the classification network it should still not be predicted. In such cases, the network’s uncertainty of the prediction can be used in order to decide whether a safe prediction can be made or not. This is called uncertainty quantification.
This thesis includes four papers covering different steps towards implementing a breast POCUS classification network. The first paper is focusing on classification of standard US images. This is a first step, showing the potential of using DL for breast cancer classification. In the second paper the classification network is modified to classify POCUS data. In order to expand the POCUS data, augmentation and CycleGAN are used. The third paper cover OOD detection and methods such as energy score and deep ensembles are studied. Finally, uncertainty quantification is covered in the fourth paper. Excluding samples with high uncertainty did improve the classification result. The papers included in this thesis show that there is potential of using DL in a trustworthy way for breast cancer detection in LMICs. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/0d77168d-83c0-44e7-9871-7d4fe2cee093
- author
- Karlsson, Jennie LU
- supervisor
-
- Anders Heyden LU
- Niels Christian Overgaard LU
- Ida Arvidsson LU
- Kalle Åström LU
- Kristina Lång LU
- organization
- publishing date
- 2024
- type
- Thesis
- publication status
- published
- subject
- keywords
- Medical Image Analysis, Breast Cancer, Point-of-Care Ultrasound, Convolutional Neural Networks, Deep Learning
- in
- Licentiate theses in mathematical sciences
- volume
- 2024
- issue
- 2
- publisher
- Lund University / Centre for Mathematical Sciences /LTH
- ISSN
- 1404-028X
- ISBN
- 978-91-8104-157-6
- 978-91-8104-158-3
- language
- English
- LU publication?
- yes
- id
- 0d77168d-83c0-44e7-9871-7d4fe2cee093
- date added to LUP
- 2024-09-16 11:06:07
- date last changed
- 2024-10-15 09:34:24
@misc{0d77168d-83c0-44e7-9871-7d4fe2cee093,
abstract = {{Breast cancer has a profound affect on society. The survival for women in low- and middle-income countries (LMICs) is poor compared to in high-income countries (HICs). The lack of timely diagnosis is one of the main factors contributing to the poor outcomes for women in LMICs. Point-of-care ultrasound (POCUS) combined with a deep learning (DL) classification network could potentially be a suitable support tool for breast cancer detection in LMICs.<br/><br/>There are different ways of designing a classification network. Convolutional neural networks (CNNs) are widely used for image classification tasks and are becoming regular in medical applications. To train a DL classification network data is needed. Medical data can be difficult to acquire due to many different reasons, one of them being ethical approvals. The availability of breast POCUS data is limited. However, the access to standard ultrasound (US) images is greater. There are different methods to expand a data set, a very common one is the use of data augmentation. Another interesting method is the cycle-consistent adversarial network (CycleGAN). This network is trained to transform an image of one domain into another domain. Thus, standard US images could be transformed into the domain of POCUS imaging, generating more POCUS samples without the need of collecting the POCUS images in the clinic. Further, it is crucial to assure the classification network to be trustworthy. Images which are out-of-distribution (OOD) should be detected and no prediction by the network should be made. OOD samples in breast US imaging includes; images of poor quality, images capturing other structures than breast tissue and images showing rare lesions. For the third case, rare lesions, the sample is in-distribution (ID). However, since the lesion is rare and might not be covered within the knowledge of the classification network it should still not be predicted. In such cases, the network’s uncertainty of the prediction can be used in order to decide whether a safe prediction can be made or not. This is called uncertainty quantification.<br/><br/>This thesis includes four papers covering different steps towards implementing a breast POCUS classification network. The first paper is focusing on classification of standard US images. This is a first step, showing the potential of using DL for breast cancer classification. In the second paper the classification network is modified to classify POCUS data. In order to expand the POCUS data, augmentation and CycleGAN are used. The third paper cover OOD detection and methods such as energy score and deep ensembles are studied. Finally, uncertainty quantification is covered in the fourth paper. Excluding samples with high uncertainty did improve the classification result. The papers included in this thesis show that there is potential of using DL in a trustworthy way for breast cancer detection in LMICs.}},
author = {{Karlsson, Jennie}},
isbn = {{978-91-8104-157-6}},
issn = {{1404-028X}},
keywords = {{Medical Image Analysis; Breast Cancer; Point-of-Care Ultrasound; Convolutional Neural Networks; Deep Learning}},
language = {{eng}},
note = {{Licentiate Thesis}},
number = {{2}},
publisher = {{Lund University / Centre for Mathematical Sciences /LTH}},
series = {{Licentiate theses in mathematical sciences}},
title = {{Deep Learning for Breast Cancer Detection in Ultrasound Imaging : Classification, Managing Data Scarcity and Detecting Out-of-Distribution Samples}},
url = {{https://lup.lub.lu.se/search/files/195249399/Lic_avhandling_Jennie_Karlsson_LUCRIS.pdf}},
volume = {{2024}},
year = {{2024}},
}