Segmentation of Lung Tumours in Positron Emission Tomography Scans: A Machine Learning Approach

Abstract

Abstract. Lung cancer represents the most deadly type of malignancy. In this work we propose a machine learning approach to segmenting lung tumours in Positron Emission Tomography (PET) scans in order to provide a radiation therapist with a "second reader" opinion about the tumour location. For each PET slice, our system extracts a set of attributes, passes them to a trained Support Vector Machine (SVM), and returns the optimal threshold value for distinguishing tumour from healthy voxels in that particular slice. We use this technique to analyse four different PET/CT 3D studies. The system produced fairly accurate segmentation, with Jaccard and Dice's similarity coefficients between 0.82 and 0.98 (the areas outlined by the returned thresholds vs. the ones outlined by the reference thresholds). Besides the high level of geometric similarity, a significant correlation between the returned and the reference thresholds also indicates that during the training phase, the learning algorithm effectively acquired the dependency between the extracted attributes and optimal thresholds.

A review on segmentation of positron emission tomography images

“…Examples of supervised and unsupervised methods used in PET segmentation include k -nearest neighbor ( k -NN) [111, 112, 113], support vector machine (SVM) [132, 133], Fuzzy C - Means (FCM) [116], artificial neural network (ANN) [111], and more recently Affinity Propagation (AP) [134, 135] and spectral clustering [119]. Clustering methods aim at gathering items with similar properties (i.e., intensity values, spatial location, etc) into local groups.…”

Thirty years of artificial intelligence in medicine (AIME) conferences: A review of research themes

“…For instance, Blonda et al [135] used a self-organizing map for segmentation and a multilayer perceptron for classification of brain MRI. Alternatively, Kerhet et al [136] used support vector machines to segment lung tumors in positron emission tomography (PET) scans.…”

Joint solution for PET image segmentation, denoising, and partial volume correction

“…These methods are easy to implement but difficult to generalize due to lack of information on local intensity distribution and sub-optimal thresholding levels. Alternatively, more sophisticated approaches have been proposed such as machine learning techniques that exploit local appearance: Gaussian mixture model and supervised/unsupervised clustering methods (Foster et al (2014b); Kerhet et al (2009)) belong to this category. For defining a specific region of interest, region growing (Li et al (2008)), level-set (Hsu et al (2008)), and graph-cut (Bagci et al (2011)) are some of the most popular region and boundary-based methods.…”