A FRAMEWORK FOR AUTOMATED TUMOR DETECTION IN THORACIC FDG PET IMAGES USING TEXTURE-BASED FEATURES G.V. Saradhi 1 , G. Gopalakrishnan 2 , A.S. Roy 2 , R. Mullick 2 , R. Manjeshwar 3 , K. Thielemans 4 , U. Patil 5 1 Computing and Decision Sciences Lab, GE Global Research, Bangalore, India 2 Imaging Technologies, GE Global Research, Bangalore, India 3 Functional Imaging Lab, GE Global Research, Niskayuna, USA 4 Hammersmith Imanet, London, UK 5 Department of Radiology, Manipal Hospital, Bangalore, India ABSTRACT This paper proposes a novel framework for tumor detection in Positron Emission Tomography (PET) images. A set of 8 second-order texture features obtained from the gray level co- occurrence matrix (GLCM) across 26 offsets, together with uptake value was used to construct a feature vector at each voxel in the data. Volume of Interest (VOI) samples from 42 images (7 patients with 6 gates each), marked by a radi- ologist, representing 5 distinct anatomy types and pathology were used to train a logit boost classifier. A ten-fold cross- validation showed a true positive rate of 96%and a false posi- tive rate of 8% for tumor classification. The test dataset con- sisted of 50 × 50 × 40 representative VOIs from gated PET images of 3 patients. The classifier was run on the test data, followed by an SUV-based thresholding and elimination of noise using connected component analysis. The method de- tected 10/12 (83%) tumors while detecting an average of 20 false positive structures. Index Terms— Positron Emission Tomography (PET), tumor classification, gray-level co-occurrence matrix, texture, logit boost 1. INTRODUCTION Tumor detection in PET images is useful in clinical appli- cations such as staging and therapy planning, and medical image analysis techniques like constraint-based non-rigid registration [1]. There are multiple automatic tumor detection techniques that have been applied successfully to the CT, Ultrasound, MR and PET modalities. Detection of tumors on PET images is a challenging problem, due to its limited spatial resolution and low signal-to-noise ratio. The prob- lem of tumor detection on PET images has been addressed in literature in various contexts. While some attempt a full- volume segmentation of a PET image, others address tumor ROI delineation. They also differ from each other in terms of the datasets on which the techniques are proposed. Kanakatte et al. [2] present a pilot study on automatic lung-tumor seg- mentation using standard uptake values (SUVs) making use of the high metabolization in the tumor. They demonstrate their results on 44 slices containing tumor and heart. A three stage approach consisting of preprocessing, segmentation and asymmetry detection was proposed by Chen et al. [3]. The authors use dynamic PET frames and compress them to four slices using principal component analysis (PCA). Further they use a graph-theoretic energy-minimization approach for segmentation, followed by an asymmetric feature detection to isolate pathological lesions in neuro images. Results are demonstrated on simulated and clinical datasets. Wong et al. [4] have used cluster analysis to replace manual ROI delin- eation in dynamic PET images. The authors use shape and magnitude of tissue Time Activity Curves (TACs) to classify them into a smaller number of distinct characteristic classes that are mutually exclusive so that the tissue TACs within a cluster are similar to one another but are dissimilar to those drawn from other clusters. They validate their technique on a simulated phantom data and assess its performance on a real dynamic PET data. Huang et al. [5] build statistical models on mean positron emission rate, raw sinogram data and the reconstructed image, and use them to derive the test criteria for maximum likelihood ratio test and a composite hypothesis test. They demonstrate detection of lesions (with a probability of 0.9) of size 15mm with lesion-to-background contrast 1.1:1. Montgomery et al. [6] demonstrate au- tomatic PET volume segmentation using Markov Random Field Models initialized by marginal segmentation, to char- acterize the spatial relationship between neighboring voxels. A spherical mean shift-based segmentation technique using a user defined seed point is proposed by [7] et al. The relation- ship between source-to-background ratio and the iso-activity level is used for segmentation of PET volumes by Daisne et al [8]. In a recent work, Naqa et al. [9] explore feature-based approach for predicting cancer treatment outcomes. They use intensity-volume histogram metrics and shape and tex- ture features extracted from PET images, to predict patient 97 978-1-4244-3932-4/09/$25.00 ©2009 IEEE ISBI 2009