- Title Information
- Title
- Segmentation of Glioblastoma Images Using Ensembles of Convolutional Neural Networks
- Type of Resource
- text
- Name:
Personal
- Name Part
- Montagut, Julian
- Role
- Role Term:
Text
- creator
- Name:
Personal
- Name Part
- Neretti, Nicola
- Role
- Role Term:
Text
- Reader
- Name:
Personal
- Name Part
- Srivastava, Vikas
- Role
- Role Term:
Text
- Reader
- Name:
Personal
- Name Part
- Lee, Jonghwan
- Role
- Role Term:
Text
- Advisor
- Name:
Corporate
- Name Part
- Brown University. Biology and Medicine: Biomedical Engineering
- Role
- Role Term:
Text
- sponsor
- Origin Information
- Copyright Date
- 2020
- Physical Description
- Extent
- ix, 91 p.
- digitalOrigin
- born digital
- Note:
thesis
- Thesis (Sc. M.)--Brown University, 2020
- Genre (aat)
- theses
- Abstract
- Brain cancer remains a pervasive public health issue. The use of CNN’s to segment brain tumor images has gained significant traction since 2012; however, most such studies utilized MRI data for these purposes, rather than stained histopathological samples used for initial diagnosis. Segmentation of these samples could provide valuable insight alongside other imaging modalities and methods. Therefore, we present patch-based segmentation of histopathological glioblastoma images using CNN’s, examining the results both in terms of AUC and MAP. After expert labeling of four primary H&E-stained histology images, we divided them into patches of size 30 x 30. We also performed data augmentation to increase the number of patches for effective training. We then grouped these patches into training, validation and testing sets. We built a modified AlexNet architecture by using Tensorflow. First, we determined optimal learning rates for AUC and MAP. Second, we used the CNN’s to generate heatmaps of the tumor probabilities of all patches in the testing set. We then further investigated the effects of heatmap averaging, median filtering and Gaussian filtering on the final predictions and identified the optimal settings over all test images. Finally, we optimized and applied thresholds to convert our probability heatmaps to classification mappings and compared the results to the reference labels. Overall, our trained CNN’s produced satisfactory segmentations, and the heatmap averaging and greater degrees of filtering tended to entail higher AUC and MAP values; however, the performance varied with which primary image was designated for testing. Furthermore, universally applying a single threshold across all test images degraded the quality of the segmentations. Selecting the optimal threshold using precision-recall curves instead of ROC curves resulted in fewer false positives, but more false negatives. The results in the MAP study were also more scattered than their AUC counterparts. Therefore, we recommend using MAP only where critical to control the frequency of false positives.
- Subject
- Topic
- Machine Learning
- Subject (fast)
(authorityURI="http://id.worldcat.org/fast", valueURI="http://id.worldcat.org/fast/00957675")
- Topic
- Histology, Pathological
- Subject (fast)
(authorityURI="http://id.worldcat.org/fast", valueURI="http://id.worldcat.org/fast/00817247")
- Topic
- Artificial intelligence
- Subject
- Topic
- Glioblastoma
- Subject
- Topic
- Deep Learning
- Subject
- Topic
- Image Segmentation
- Subject (fast)
(authorityURI="http://id.worldcat.org/fast", valueURI="http://id.worldcat.org/fast/00837592")
- Topic
- Brain--Cancer
- Language
- Language Term (ISO639-2B)
- English
- Record Information
- Record Content Source (marcorg)
- RPB
- Record Creation Date
(encoding="iso8601")
- 20200720
- Access Condition:
rights statement
(href="http://rightsstatements.org/vocab/InC/1.0/")
- In Copyright
- Access Condition:
restriction on access
- Collection is open for research.
- Identifier:
DOI
- 10.26300/dcax-ve55