Please use this identifier to cite or link to this item:
|Title:||Deep learning models for survival analysis on histology images|
|Keywords:||Convolutional neural network;survival analysis;histology;region of interest (ROI);high power fields (HPFS);DensNet 121;VGG-19;xception;inception-V3;inception-ResNet-V2|
|Abstract:||In this study, Histology images of the patients were used as input to the Convolutional Neural Network (CNN) models to predict the risk of the patients with brain tumour. Motivation for this study was to highlight the emerging role of deep learning in the field of precision medicine and suggest an expanding utility for computational analysis of histology in the future practice of pathology. The Region of Interest (ROI) of the histology images (1024x1024 pixels) of brain tumour of 769 patients was used to build a machine learning model which can predict the hazard ratio (is frequently interpreted as risk ratio) of the patients and survival time of the patient. Five diverse CNN models have been trained namely, DensNet121, VGG-19, Xception, Inception-V3 and Inception-ResNet-V2. The loss function of Cox-proportional hazard has been used to fit a survival model. For each CNN survival model, a 15-fold cross validation was implemented on the training data. The image data for 769 patients was split into 80% as training with 616 patients and 20% for testing consisting of 153 patients. There are 1239 ROI images for 616 patients and 266 ROI images of 153 patients. For each cross validation, the CNN model was trained for 100 epochs. For the test data (153 patients), 9 High Power Fields (HPFs) (256x256 pixels) were sampled from each ROI, and a risk is predicted for each field. The Median HPF risk is calculated in each ROI and the second highest value among all ROIs was selected as the patient risk. The predicted risk is calculated by taking the average of predicted risk for the last 5 epochs. The median concordance index (CI) and integrated brier score (IBS) is evaluated from the predicted risk and are compared for each of the models. Wilcoxon Ranksum test is used to test the null hypothesis that the CI and IBS of five CNN models was significantly different for 5% significance level. From the five CNN structures, DensNet121 had the best performance, which was followed by VGG19, InceptionResNet-V2, InceptionV3 and Xception, respectively.|
|Appears in Collections:||Computational Sciences - Master's theses|
Files in This Item:
|Thesis FINAL - Hetang Patel - 03-June-21.docx.pdf||5.76 MB||Adobe PDF||View/Open|
Items in LU|ZONE|UL are protected by copyright, with all rights reserved, unless otherwise indicated.