A new method of computer-aided feature identification for lesion detection in PET-FDG dynamic study

Computer-aided feature identification in PET-FDG dynamic data is crucial to assist visual inspection for small lesion detection. It has been shown that the kinetic features of time activity curves (TAG) have the following properties: (a) linearly representable by a set of exponential functions, (b) physiologically distinguishable as lesion and normal tissue subspaces, (c) readily incorporable to a matched subspace detector for lesion detection. To identify the TAC subspace features, the least square error (LSE) method is often used to determine the lesion and normal tissue subspaces respectively The subspaces resulted from the LSE are optimum in terms of the fidelity to the observed data, but they may suffer from a lack of the separability between subspaces. Here, a new method is proposed to maximize the distance (separability) between lesion and normal tissue subspaces under the constraint that the LSE (fidelity) of the estimated and observed TACs is less than a given value. Such identified subspaces are incorporated into a matched subspace detector for lesion detection. Results showed that the subspaces identified by the proposed method from known lesion and normal tissues mostly preserve the TAC kinetic features and increase the contrast of the small lesion to normal tissues compared to the LSE-only method