To assess the performance of each filter, many metrics were computed, including contrast-to-noise ratio (CNR), comparative signal-to-noise ratio (cSNR), mean squared error (MSE), and structural similarity (SSIM). These metrics are briefly explained at the bottom of this page.
A table of the average results of 30 tissue-mimicking phantoms can be seen here, with the above metrics highlighted. Each block section is of a filter type (or the eREC-FC process alone) and each column shows results from a window size (for median, homogeneous mask area, and Lee) or number of iterations (for geometric and SRAD).
Important things to note from the data:
- SRAD can improve CNR by an average 560% over CP.
- SRAD can decrease the noise (speckle) and increase the signal
information by an average 70% over eREC-FC.
- SRAD more resembles the original CP image than eREC-FC does,
based on MSE, and has the best structural resemblance to CP of all
the filtering techniques.
CNR - quantifies the degree to which a region stands out from another region. Larger numbers show better results.
cSNR - tells the degree to which noise (speckle) was suppressed in one image over the next. All images are compared to CP. Larger numbers show better results.
MSE - shows the average absolute difference between two images. All images are compared to CP. A smaller number is desired.
SSIM - denotes how well two images resemble each other, where 1 represents high similarity and -1 means no resemblance.
The other metrics used are extensions of the four above and are defined in the final project paper, in section III-B.
Metrics used to assess filter performance.