Revise Compactness 1 Formulas #218
Conversation
Flip exponent (2/3 -> 3/2) of Compactness 1 feature formula. The previous exponent was most likely an error, as this resulted in a formula that was not dimensionless, as the description indicates. The formula with the new exponent is dimensionless. Compactness1, Compactness2, Sphericity and Spherical disproportion are mathematically correlated. Update the documentation to reflect this (provide equations specifying the correlation between the different features). Also disable Compactness 1 and Compactness 2 in the default/example parameter file in the bin folder to emphasize the redundancy of these features. Updated the formulas and documentation thereof to have simplified formulas where possible.
Remove the Roundness Feature, as this expresses the same formula as Sphericity, but is calculated by Simple ITK's LabelShapeStatisticsImageFilter (LSSIF). It is documented here: http://www.insight-journal.org/browse/publication/176. The calculated values differ however, but this is due to the fact that SimpleITK calculates the surface area differently. The surface area calculated by SimpleITK can be found by calling LSSIF.GetPerimeter(). Calculation of Roundness and Surface Area is not done using SimpleITK, as the calculated surface area can be smaller than the surface area of a sphere with the same volume as the ROI (and therefore causes Roundness > 1). Example in lung1 testcase: * SA (pyradiomics) 782.24 * SA (SimpleITK) 576.06 * SA of a sphere with volume 1361.20: 593.97
JoostJM
force-pushed
the
revise-compactness
branch
from
1cf1dd5 to
7290f88
Compare
Update baseline to reflect change in Compactness 1 formula and the removal of the Roundness feature.
|
LGTM - sounds like a good fix. |
Hi @JoostJM, I'm a radiomics researcher attempting to validate a model proposed by your associate Hugo Aerts in his landmark paper "Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach". I'm reaching out to you here because I have a question about the calculation of compactness1 as it pertains to this work and the validation paper "External validation of a prognostic CT-based radiomic signature in oropharyngeal squamous cell carcinoma". In both papers, the supplementary materials describe compactness1 as having been calculated with an exponent of 2/3 on the area. I understand the value of changing the exponent to 3/2 such that the metric is dimensionless as its description states. However, I am curious to know how this feature was calculated in these papers for the purposes of a successful validation. Specifically, I want to know if they used an exponent of 2/3 as reported in the supplementary materials and this was then changed in this update of pyradiomics or if this was a typo that was then implemented in pyradiomics and needed changing afterwards. Any guidance would be appreciated. Thanks in advance |
Flip exponent (2/3 -> 3/2) of Compactness 1 feature formula. The previous exponent was most likely an error, as this resulted in a formula that was not dimensionless, as the description indicates. The formula with the new exponent is dimensionless.
Compactness1, Compactness2, Sphericity and Spherical disproportion are mathematically correlated. Update the documentation to reflect this (provide equations specifying the correlation between the different features). Also disable Compactness 1 and Compactness 2 in the default/example parameter file in the bin folder to emphasize the redundancy of these features.
Updated the formulas and documentation thereof to have simplified formulas where possible.