Diagnosis of Parkinsonian disorders using a channelized Hotelling observer model: proof of principle

• Published in: Medical physics (Lancaster) Vol. 34; no. 10; p. 3987
• Main Authors: Bal, H, Bal, G, Acton, P D
• Format: Journal Article
• Language: English
• Published: United States 01.10.2007
• Subjects: Algorithms; Automation; Brain - diagnostic imaging; Brain - pathology; Computer Simulation; Dopamine Plasma Membrane Transport Proteins - metabolism; Humans; Models, Biological; Models, Statistical; Models, Theoretical; Normal Distribution; Parkinsonian Disorders - diagnosis; Parkinsonian Disorders - diagnostic imaging; Phantoms, Imaging; Positron-Emission Tomography - methods; Radiography; Reproducibility of Results; Tomography, Emission-Computed, Single-Photon - methods
• ISSN: 0094-2405; 2473-4209
• DOI: 10.1118/1.2776250

Imaging dopamine transporters using PET and SPECT probes is a powerful technique for the early diagnosis of Parkinsonian disorders. In order to perform automated accurate diagnosis of these diseases, a channelized Hotelling observer (CHO) based model was developed and evaluated using the SPECT tracer [Tc-99m]TRODAT-1. Computer simulations were performed using a digitized striatal phantom to characterize early stages of the disease (20 lesion-present cases with varying lesion size and contrast). Projection data, modeling the effects of attenuation and geometric response function, were obtained for each case. Statistical noise levels corresponding to those observed clinically were added to the projection data to obtain 100 noise realizations for each case. All the projection data were reconstructed, and a subset of the transaxial slices containing the striatum was summed and used for further analysis. CHO models, using the Laguerre-Gaussian functions as channels, were designed for two cases: (1) By training the model using individual lesion-present samples and (2) by training the model using pooled lesion-present samples. A decision threshold obtained for each CHO model was used to classify the study population (n = 40). It was observed that individual lesion trained CHO models gave high diagnostic accuracy for lesions that were larger than those used to train the model and vice-versa. On the other hand, the pooled CHO model was found to give a high diagnostic accuracy for all the lesion cases (average diagnostic accuracy = 0.95 +/- 0.07; p < 0.0001 Fisher's exact test). Based on our results, we conclude that a CHO model has the potential to provide early and accurate diagnosis of Parkinsonian disorders, thereby improving patient management.