Murdoch University Research Repository

Welcome to the Murdoch University Research Repository

The Murdoch University Research Repository is an open access digital collection of research
created by Murdoch University staff, researchers and postgraduate students.

Learn more

Deep learning-based scale-invariant cancer detection from whole slide image

Zaidi, Fatima (2021) Deep learning-based scale-invariant cancer detection from whole slide image. Masters by Research thesis, Murdoch University.

[img]
Preview
PDF - Whole Thesis
Download (5MB) | Preview

Abstract

Convential cancer diagnosis methods from whole slide images (WSI) train a deep Convolutional Neural Network (CNN) to make patch level predictions, and then aggregate the image-level predictions to classify a tumour as either benign or malignant. To classify a patch, the CNN extracts features through convolutional layers and then process the feature maps using fully connected layers. The size of the filters used in the convolutional layers defines the receptive field of the network. Small filters are computationally efficient but do not capture a large context. On the other hand, large filters allow learning features that capture a larger context but are very expensive both in terms of computational time and memory requirements. This paper focuses on two main challenges. The First one is how to incorporate a large context while minimizing the computational overhead. The second one is that the cancerous cells can be of arbitrary size, and thus any detection and recognition approach should be scale-invariant. We introduce the Dilated SPP VGG-16 network with different dilation rates applied to every block of the VGG-16 network. The proposed dilated SPP VGG-16 architecture allows to increase the receptive field of the network without increasing the filter size or the depth of the network, and thus are very efficient to train. It also enables the multiscale analysis without changing the architecture of the network and retraining. We tested the proposed approach on the publicly available Camelyon17 dataset. Our experiments show that the proposed CNN achieves comparable or better accuracy than a conventional deep learning method, but with significantly less computational time and memory requirements.

Item Type: Thesis (Masters by Research)
Murdoch Affiliation(s): IT, Media and Communications
Supervisor(s): Laga, Hamid, Sohel, Ferdous and Koutsakis, Polychronis
URI: http://researchrepository.murdoch.edu.au/id/eprint/63326
Item Control Page Item Control Page

Downloads

Downloads per month over past year