Rotational Invariant Convolutional Neural Networks for Prostate Cancer Classification
(2017) In Master's Theses in Mathematical Sciences FMA820 20171Mathematics (Faculty of Engineering)
- Abstract
- Prostate cancer was in 2012 the second most common type of cancer for males globally. To be able to treat prostate cancer in the most effective way it is important to know how aggressive the cancer is. This aggressiveness is graded using the Gleason score. The diagnosis is done by pathologists inspecting prostate biopsies, but the advancement of pattern recognition using Convolutional Neural Networks (CNN) has made it interesting to try to automate this process.
The data in this thesis is microscopy images of prostatic tissue. These images are rotation-invariant, meaning that they have the same Gleason grade no matter which angle they are inspected at. The goal of this thesis is to investigate the possibility of exploiting this... (More) - Prostate cancer was in 2012 the second most common type of cancer for males globally. To be able to treat prostate cancer in the most effective way it is important to know how aggressive the cancer is. This aggressiveness is graded using the Gleason score. The diagnosis is done by pathologists inspecting prostate biopsies, but the advancement of pattern recognition using Convolutional Neural Networks (CNN) has made it interesting to try to automate this process.
The data in this thesis is microscopy images of prostatic tissue. These images are rotation-invariant, meaning that they have the same Gleason grade no matter which angle they are inspected at. The goal of this thesis is to investigate the possibility of exploiting this rotation invariance to create a rotation invariant Convolutional Neural Network for automatic classification.
The rotation invariant CNNs include a rotation of filters, which makes interpolation an important aspect to investigate. This thesis does that by designing many different CNNs in a general way that have different sizes of the filters that are to be rotated.
The resulting CNNs show that the filter size does indeed matter, with the smallest rotated filters that trained the network well were of size 17x17. The lowest resulting error rate for classification on both the training and validation data was 6.3%. The lowest error rate for classification on just the validation data was 16.7%, however the validation data consisted of only 24 images. The conclusion from this was that making the CNN rotation invariant can be of some interest, and could be investigated further by optimizing networks for a certain size of rotating filters. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/8926862
- author
- Ekelund, Joel LU
- supervisor
-
- Anders Heyden LU
- Ida Arvidsson LU
- organization
- course
- FMA820 20171
- year
- 2017
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Convolutional Neural Networks, rotation invariance, deep learning, automated Gleason grading
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMA-3334-2017
- ISSN
- 1404-6342
- other publication id
- 2017:E64
- language
- English
- id
- 8926862
- date added to LUP
- 2017-12-01 15:28:34
- date last changed
- 2017-12-01 15:28:34
@misc{8926862, abstract = {{Prostate cancer was in 2012 the second most common type of cancer for males globally. To be able to treat prostate cancer in the most effective way it is important to know how aggressive the cancer is. This aggressiveness is graded using the Gleason score. The diagnosis is done by pathologists inspecting prostate biopsies, but the advancement of pattern recognition using Convolutional Neural Networks (CNN) has made it interesting to try to automate this process. The data in this thesis is microscopy images of prostatic tissue. These images are rotation-invariant, meaning that they have the same Gleason grade no matter which angle they are inspected at. The goal of this thesis is to investigate the possibility of exploiting this rotation invariance to create a rotation invariant Convolutional Neural Network for automatic classification. The rotation invariant CNNs include a rotation of filters, which makes interpolation an important aspect to investigate. This thesis does that by designing many different CNNs in a general way that have different sizes of the filters that are to be rotated. The resulting CNNs show that the filter size does indeed matter, with the smallest rotated filters that trained the network well were of size 17x17. The lowest resulting error rate for classification on both the training and validation data was 6.3%. The lowest error rate for classification on just the validation data was 16.7%, however the validation data consisted of only 24 images. The conclusion from this was that making the CNN rotation invariant can be of some interest, and could be investigated further by optimizing networks for a certain size of rotating filters.}}, author = {{Ekelund, Joel}}, issn = {{1404-6342}}, language = {{eng}}, note = {{Student Paper}}, series = {{Master's Theses in Mathematical Sciences}}, title = {{Rotational Invariant Convolutional Neural Networks for Prostate Cancer Classification}}, year = {{2017}}, }