Automatic detection of active nodules in 3D PET oncology imaging using the Hotelling Observer and the Support Vector Machines: A comparison study

摘要

Positron Emission Tomography (PET) using fluorine-18-deoxyglucose (FDG) has become an increasingly recommended tool in clinical whole-body oncology imaging for detection, diagnosis and follow up of many cancers. One way to improve the diagnostic utility of PET oncology imaging is to assist the physicians facing the difficult cases of residual or low contrast lesions. This study aims at proposing and comparing two methods that perform guided detection of abnormal foci in PET based on the classification theory of Computer Aided Detection (CAD) systems. The first original method is based on the linear Hotelling Observer (HO), mostly used for image quality assessment. The second method uses a more classical non linear classifier, the Support Vector Machine (SVM), which has never been applied to lesion detection task for 3D whole-body PET imaging. The image feature sets that serve as input data for both classifiers are similar and consist of the coefficients of an undecimated wavelet transform. Detection performances of both classifiers are compared based on a simulated whole-body PET image database consisting of 250 images containing 1750 lesions for training and 25 images with 175 lesions for testing. An optimization study is performed for each classifier separately to select the best combination of parameters including the level of wavelet decomposition and the characteristics of the training database. The discriminatory power of the feature vector is also evaluated through the implementation of a Genetic Algorithm (GA). A preliminary post-processing based on the majority voting based combination of both classifiers allows reducing the number of false positive clusters per image (FPI). The CAD system including false positives reduction indicates promising classification performances with couples sensitivity/FPI of 80%/25 for the lungs, 81%/14 for the liver and 70%/21 for soft tissue considering the best combination of parameters.