ROC analysis of an automated program for analyzing striatal uptake of DaTscan SPECT images: Comparison to visual reads

Objectives A fully automated objective striatal analysis program was applied to dopamine transporter DaTscan images acquired from subjects with suspected, early, and advanced symptoms of Parkinsonism. The Striatal Binding Ratios (SBR) of the lowest putamen uptake was computed and ROC analysis was applied to a group of 94 subjects to determine the best quantitative discriminator for the scans.

Methods A total of 94 DaTscan SPECT scans were acquired from subjects referred to our clinical imaging department and reconstructed using our standard brain protocol. Visual reads of the patients were conducted by two of our trained clinical faculty; when the two reads were discordant, a third independent reader was used to obtain majority consensus. The same images were analyzed using the executable program (Objective Striatal Analysis, OSA [J. Nucl. Med. 2007 Jun; 48(6):857-64]) which locates the striatal and occipital structures, locates the caudate and putamen, and calculates the background subtracted striatal binding ratio (SBR). The lowest SBR of the two putamen regions was taken as the quantitative report. A group of 33 normal and 61 patients (whose scans were read as showing dopaminergic deficit by the readers) were analyzed to generate ROC curves.

Results The ROC analysis, which used the visual interpretation from the readers as the gold standard, yielded an AUC = 0.976 when using 2.08 as the threshold SBR value for the lowest putamen. The sensitivity and specificity of the automated quantitative analysis was 93% and 97% respectively.

Conclusions The OSA program delivers SBR quantitative values which have a high sensitivity and specificity compared to visual reads conducted by trained nuclear medicine physicians. Such a program could be a helpful aid for readers not yet experienced in reading DaTscan SPECT images, and could be used to help in assessing possible disease progression over relatively short periods of time (months).

Research Support Work conducted under support of NIH grant 5R42NS05547