Author: Kenneth A. Fetterly 👨🔬
Affiliation: Mayo Clinic 🌍
Purpose: Among the limitations of channelized Hotelling T2 type model observers (CHO) applied to medical imaging systems is that they reduce 2D image detail to a singular value and are only applicable to symmetric test objects. The purpose of this work was to develop a statistical image analysis method which preserves the spatial detail of sample images, thereby making it applicable to arbitrary test objects.
Methods: Unprocessed test signal present and absent (n=1200) angiography images were used. Stationary test objects included iodine disks, pediatric implantable vascular plugs, and guidewires on a uniform background. The pixel-specific contribution (t2) to SNR2 by Hotelling T2 was estimated by two methods. For both methods, t2 was calculated from a small square patch of pixels centered on the single pixel tested. Patch size was determined a-priori by examination of the inter-pixel noise correlation distance. The methods differed in how the patch covariance matrix was used to estimate t2. Pixel-specific t2 within a ROI including a test object was calculated in a raster scan manner, resulting in 2D maps of t2. The sum of t2 over ROI pixels provided an estimate of T2.
Results: Estimates of d’=sqrt(T2) of iodine disks by the two methods were within ±6% of d’ estimated by an established Gabor CHO. Inspection of the t2 maps demonstrated inter-pixel t2 precision differences between the two methods, which was consistent with instability of covariance matrix inversion.
Conclusion: This work provides a foundation to estimate the contribution of individual pixels to Hotelling T2. The methods were accurate compared to a Gabor function CHO and practical to implement. Future work will explore the utility of the t2 maps to describe imaging system performance and human impression of image quality.