Advanced

Classification of Prognosis in Breast Cancer Patients from AMCL Analysis using Machine Learning Techniques

Ruuskanen, Johan and Andersson, Ola (2017) FMS820 20171
Mathematical Statistics
Abstract
Predicting the development of distant metastasis for breast cancer patients is of high importance for both the patient and the medical staff. The current best method for prediction is the use of handcrafted histopatological features. The aim with this study is to explore how well Additive Multiple Labelling Cytochemistry (AMLC) stained core biopsy images can predict the development of distant metastasis. For this, two cohorts of a total of 488 patients are investigated, each patient having 1-4 images of AMLC stained core biopsies (of size 2 mm) from the tumour area and AMLC features extracted from those images. Each patient is also supplied with the handcrafted histopatolohical features for reference. Both the images and numerical AMLC... (More)
Predicting the development of distant metastasis for breast cancer patients is of high importance for both the patient and the medical staff. The current best method for prediction is the use of handcrafted histopatological features. The aim with this study is to explore how well Additive Multiple Labelling Cytochemistry (AMLC) stained core biopsy images can predict the development of distant metastasis. For this, two cohorts of a total of 488 patients are investigated, each patient having 1-4 images of AMLC stained core biopsies (of size 2 mm) from the tumour area and AMLC features extracted from those images. Each patient is also supplied with the handcrafted histopatolohical features for reference. Both the images and numerical AMLC features extracted from the images contain information on the immunological response of the patient which in turn has been shown to have potential of good predictive ability. The images were analyzed using convolutional neural networks and the AMLC features with a support vector machine, random forest and linear discriminant analysis classifiers. We show that the convolutional neural networks and the numerical classifiers achieve similar performance with an area under the receiver operating characteristic curve (AUC) value of approximately 0.60, which is worse than the result achieved on the histopathological features, which gave an AUC value of approximately 0.74. Further, we show that no difference in prediction can be observed with using AMLC images containing immunological features as compared to AMLC images not containing any immunological features. Finally, we show that an ensemble of the classifiers for the AMLC features and the images gives no significant boost to performance in terms of AUC value. Data from 488 patients were used in the study; the results indicate that the sample size was too small to capture the variance present between patients. Furthermore the core biopsies were most likely too small to capture the behaviour of the tumour. For future studies, it is recommended to increase the number of patients and the size of the core biopsies. (Less)
Please use this url to cite or link to this publication:
author
Ruuskanen, Johan and Andersson, Ola
supervisor
organization
course
FMS820 20171
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Breast Cancer, AMLC, Metastasis Prediction, Convolutional Neural Network, Machine Learning.
language
English
id
8924971
date added to LUP
2017-09-07 11:21:35
date last changed
2017-09-07 11:21:35
@misc{8924971,
  abstract     = {Predicting the development of distant metastasis for breast cancer patients is of high importance for both the patient and the medical staff. The current best method for prediction is the use of handcrafted histopatological features. The aim with this study is to explore how well Additive Multiple Labelling Cytochemistry (AMLC) stained core biopsy images can predict the development of distant metastasis. For this, two cohorts of a total of 488 patients are investigated, each patient having 1-4 images of AMLC stained core biopsies (of size 2 mm) from the tumour area and AMLC features extracted from those images. Each patient is also supplied with the handcrafted histopatolohical features for reference. Both the images and numerical AMLC features extracted from the images contain information on the immunological response of the patient which in turn has been shown to have potential of good predictive ability. The images were analyzed using convolutional neural networks and the AMLC features with a support vector machine, random forest and linear discriminant analysis classifiers. We show that the convolutional neural networks and the numerical classifiers achieve similar performance with an area under the receiver operating characteristic curve (AUC) value of approximately 0.60, which is worse than the result achieved on the histopathological features, which gave an AUC value of approximately 0.74. Further, we show that no difference in prediction can be observed with using AMLC images containing immunological features as compared to AMLC images not containing any immunological features. Finally, we show that an ensemble of the classifiers for the AMLC features and the images gives no significant boost to performance in terms of AUC value. Data from 488 patients were used in the study; the results indicate that the sample size was too small to capture the variance present between patients. Furthermore the core biopsies were most likely too small to capture the behaviour of the tumour. For future studies, it is recommended to increase the number of patients and the size of the core biopsies.},
  author       = {Ruuskanen, Johan and Andersson, Ola},
  keyword      = {Breast Cancer,AMLC,Metastasis Prediction,Convolutional Neural Network,Machine Learning.},
  language     = {eng},
  note         = {Student Paper},
  title        = {Classification of Prognosis in Breast Cancer Patients from AMCL Analysis using Machine Learning Techniques},
  year         = {2017},
}