Data Augmentation to Increase Multi-Site Robustness for Convolutional Neural Networks - A case study on MRI segmentation of target and organs at risk for prostate cancer
(2019) In Master's Theses in Mathematical Sciences FMSM01 20191Mathematical Statistics
- Abstract
- Organ segmentation on magnetic resonance (MR) images for dose planning on cancer patients is a time consuming process that can be automatized by using convolutional neural networks (CNNs). MR images vary greatly in characteristics across acquisition sites, which affects CNN segmentation performance. In this master thesis, augmentation methods were tested and implemented to increase the robustness of CNNs in a multi-site context. A data set of MR images from 151 prostate cancer patients from four Swedish hospitals (denoted A - D) was used, and the prostate, bladder, rectum, femoral heads, and body were segmented. Four networks were developed on 40 training subjects from a single site for two vendors using HighRes3DNet_large implemented in... (More)
- Organ segmentation on magnetic resonance (MR) images for dose planning on cancer patients is a time consuming process that can be automatized by using convolutional neural networks (CNNs). MR images vary greatly in characteristics across acquisition sites, which affects CNN segmentation performance. In this master thesis, augmentation methods were tested and implemented to increase the robustness of CNNs in a multi-site context. A data set of MR images from 151 prostate cancer patients from four Swedish hospitals (denoted A - D) was used, and the prostate, bladder, rectum, femoral heads, and body were segmented. Four networks were developed on 40 training subjects from a single site for two vendors using HighRes3DNet_large implemented in NiftyNet. Two networks were trained without augmentation, and two with. The networks were evaluated on average Dice score (DSC) over all organs, on a validation set consisting of images from all four sites. The DSC improved from 0.886 to 0.927 for the network trained on site A, and from 0.668 to 0.919 for the network trained on site C when augmentation was added. The applied augmentations were either a part of the NiftyNet framework, namely bias field and geometric augmentations including elastic deformation, or implemented by the authors, i.e., augmentations by histogram modification, adding noise, and smoothing. All augmentations were applied online with the aim to increase the variety during training of the CNN. Hence, the augmentations were not intended to mimic the non-training sites. The DSCs when using augmentation are comparable to scores for a network trained on a large multi-site data set (0.930) as well as organ segmentation variability between experts. The DSC is based on pixel overlap between the network segmentation and the ground truth, which was segmented by non-experts using available guidelines. This thesis shows the importance of augmentation for multi-site segmentation and presents useful tools as a stepping stone towards automatizing organ segmentations. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/8994188
- author
- Börnfors, Frida LU and Klint, Elisabeth
- supervisor
- organization
- course
- FMSM01 20191
- year
- 2019
- type
- H2 - Master's Degree (Two Years)
- subject
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMS-3379-2019
- ISSN
- 1404-6342
- other publication id
- 2019:E48
- language
- English
- id
- 8994188
- date added to LUP
- 2019-09-17 12:04:11
- date last changed
- 2019-09-17 12:04:11
@misc{8994188, abstract = {{Organ segmentation on magnetic resonance (MR) images for dose planning on cancer patients is a time consuming process that can be automatized by using convolutional neural networks (CNNs). MR images vary greatly in characteristics across acquisition sites, which affects CNN segmentation performance. In this master thesis, augmentation methods were tested and implemented to increase the robustness of CNNs in a multi-site context. A data set of MR images from 151 prostate cancer patients from four Swedish hospitals (denoted A - D) was used, and the prostate, bladder, rectum, femoral heads, and body were segmented. Four networks were developed on 40 training subjects from a single site for two vendors using HighRes3DNet_large implemented in NiftyNet. Two networks were trained without augmentation, and two with. The networks were evaluated on average Dice score (DSC) over all organs, on a validation set consisting of images from all four sites. The DSC improved from 0.886 to 0.927 for the network trained on site A, and from 0.668 to 0.919 for the network trained on site C when augmentation was added. The applied augmentations were either a part of the NiftyNet framework, namely bias field and geometric augmentations including elastic deformation, or implemented by the authors, i.e., augmentations by histogram modification, adding noise, and smoothing. All augmentations were applied online with the aim to increase the variety during training of the CNN. Hence, the augmentations were not intended to mimic the non-training sites. The DSCs when using augmentation are comparable to scores for a network trained on a large multi-site data set (0.930) as well as organ segmentation variability between experts. The DSC is based on pixel overlap between the network segmentation and the ground truth, which was segmented by non-experts using available guidelines. This thesis shows the importance of augmentation for multi-site segmentation and presents useful tools as a stepping stone towards automatizing organ segmentations.}}, author = {{Börnfors, Frida and Klint, Elisabeth}}, issn = {{1404-6342}}, language = {{eng}}, note = {{Student Paper}}, series = {{Master's Theses in Mathematical Sciences}}, title = {{Data Augmentation to Increase Multi-Site Robustness for Convolutional Neural Networks - A case study on MRI segmentation of target and organs at risk for prostate cancer}}, year = {{2019}}, }