Novel image processing and deep learning methods for head and neck cancer delineation from MRI data

Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2022
Thesis identifier
  • T16332
Person Identifier (Local)
  • 201580626
Qualification Level
Qualification Name
Department, School or Faculty
  • Intensity modulated radiation treatment aims to achieve accurate treatment of cancer without introducing damage and side effects to organs at risk (OAR). Development of medical imaging technique enables molecular study of cancer to provide quantitative analysis and three-dimensional visualisation to oncologists and radiotherapists for better radiation treatment planning (RTP). Conventional radiation treatment process of head and neck cancer (HNC) requires the manual delineation of gross tumour volume (GTV), abnormal (cancerous) lymph nodes (ALNs), and organs at risk. While the manual delineation subjects to inter- and intra- observer variabilities. Novel image processing and deep learning method for head and neck cancer delineation from MRI data are presented. Firstly, a head and neck ALNs segmentation pipeline including pre-processing of MRI data, knowledge-based detection, and extraction of 3D volume of ALNs is presented. Secondly, a 3D HNC delineation via improving detection and segmentation modules is presented. In these methods, T1 axial MRI slices were firstly pre-processed using contrast enhancement (CE), background noise removal, and bias field correction to improve image quality. 2D slices were then interpolated vertically to reconstruct 3D MRI volume. A knowledge based ALNs detection algorithm was proposed to use throat as key spatial landmark and use fuzzy c-mean (FCM) to classify tissues in intensity, so that find the ALNs on each slice. The 2D detection results were ensembled by a proposed majority voting scheme to give the 3D location of ALN in MRI volume. The 3D volume was finally extracted by 3D level set method (LSM) starting from the detected centre of ALN. The knowledge-based detection method achieved localisation of ALNs by transferring clinical knowledge to automatic algorithm, and by ensemble of results of multiple slices to improve confidence level of detection. This method provided objective 3D segmentation, visualisation, and quantification of ALNs from MRI data, the delineated ALN volumes were comparable (70% in DSC) to conventional manual delineation but with a lower time cost. Furthermore, a knowledge-based method for segmentation of 3D volume of HNC from MRI data was proposed. This method also has pre-processing, detection, and segmentation steps. In pre-processing stage, the raw MRI data went through CE, bias field correction, intensity standardisation (IS), and vertical interpolation to generate 3D reconstructed MRI volume with better quality. The target HNC was found by knowledge-based detection on central slice in MRI volume. Detection used FCM to classify tissues, used throat to guide spatial searching, and used localized LSM to refine the region of 2D detection. The detection gave location of 3D volume of HNC and kept spatial information in central slice. A modified 3D LSM was started from detected volume centre to extract the 3D volume of HNC. The extracted volume was smoothed by morphological filtering. The interpolation and 3D segmentation method extracted uniform smooth 3D HNC volume from 2D T1-axial MRI slices. The modified 3D LSM improved the accuracy of volume segmentation via combining spatial information to supress the false positive (FP), i.e., overestimation in segmentation. The proposed automatic 3D segmentation method achieved comparable (70% in DSC) 3D volume of HNC with manual segmentation but lower time cost. Algorithm was further developed to window-based software as useful RTP tool. Thirdly, a new DCNN for pixel-wise end-to-end segmentation of HNC from 2D T1 axial MRI slices is presented, this architecture was trained and tested on manual consensus outlined from clinicians. The network took similar structure from classical network U-Net, which included an encoder part to extract features and reduce resolution, then a decoder part to recover resolution, fuse features, and classify pixels. The proposed new DCNN improved feature extraction by using a two-pathway encoder with classical and dilated convolutional kernels to combine local and non-local information. To design this DCNN using limited HNC MRI data, data augmentation was used to help the training, depth-separable convolution was used to reduce number of parameters, cross-validation was used to avoid overfitting. The proposed DCNN improved accuracy of HNC segmentation from real MRI data by 5% compared to classical U-Net.
Advisor / supervisor
  • Soraghan, John
  • Caterina, Gaetano Di
Resource Type