Recent Changes - Search:


edit SideBar


Decoding Tumour Phenotype by noninvasive Imaging using a Quantitative Radiomics Approach

H.J. Aerts et. al. NATURE COMMUNICATIONS 5, 4006, June 2014.

The paper describes the use of image based metrics for tumour grading. This approach seems to be motivated largely by a wish to avoid the manual selection of pattern recognition (PR) variables. This avoids the researcher needing either an understanding of the feature computations, or the behaviour of the training data. A package (here MATLAB ) is used to generate a large number of candidate features as the basis of a predictive subset. Several hundred possible features are evaluated entirely empirically, using (a mere) 600 data samples. The resulting PR system uses only 4 of these. Performance is claimed to be sufficiently good to be clinically useful, some might even say impressive. However, there may be less to this approach than meets the eye.

The well known difficulties involved with selecting feature vectors using finite quantities of training data are not mentioned. The paper emphasises the large number of possible features, and supplementary materials provide some nominal mathematical description. Neither discuss the properties of the selected features themselves. I will therefore attempt to do this for you here. The features selected are;

A) The sum of squares of the grey-level intensities in the ROI.

B) A measure of "sphericity" of the ROI (expressed as a suitable polynomial ratio of surface area and volume).

C) A 3D wavelet based calculation taken (somehow) from within the ROI.

D) A 3D run-length summary of identical pixels within the ROI.

Some interesting properties of these selected features are as follows.

`A' gets bigger for larger regions and greater regional proportions of image enhancement. `B' Is larger for irregular shapes (smallest for a sphere). `C' Is dependant upon the co-ordinate system of the data (the calculation is specifically aligned to the `y' axis). The original paper and supporting documentation does not specify how a single number is constructed from a whole set of pixels in the region, nor which scale of wavelet is selected. We also do not know if the wavelet is applied only to data within the ROI (setting all external data to zero) or across the ROI boundary (thereby including surrounding material). `D' summaries something regarding the local texture of the ROI, but will be dependant upon an (arbitrary) grey-level quantization level, which is not documented in the paper. Also the specified calculation returns multiple values (for 16 possible orientations) but I can find no mention of how these are used to compute the single value needed for recognition.

The first two of these are rather obvious indicators of tumour aggression, but would probably not have been considered particularly impressive if used alone. Some of the properties of `C' and `D' do not make logical sense, and so could result in problems when applied to data outside of the original training and test sets. Indeed many of the variables input to this system have either poor statistical stability or depend inappropriately on uncontrolled properties of data. They will have been selected here due to the small quantity of data and poor methodology (i.e. the dataset does not encompass sufficient variation). The performance assessment is inappropriate, as it concentrates on showing that the prognosis distribution matches the ground truth distribution, rather than correlations of ground truth with individual prognoses.

In summary. The publication fails to provide sufficient detail to allow replication. The selected variables are either trivial, unsuitable or remarkably both. The failing of this approach seems to be that if a researcher doesn't understand the variables he feeds in to the selection process, then he will also be unable to comprehend the properties (good or bad) of those that come out. The rather weak assessment measure probably accounts for the equally good performance seen with completely different tumours.

I believe that this paper is a useful starting point for assessment of an increasing trend; the use of complex software as black boxes and an over-reliance on poorly executed empiricism. Other papers in this review list show some of the many pitfalls associated with this. Oncologists should take note, there is enough evidence to show that one favourable PR assessment does not prove an idea.

NAT 1/6/2015

Edit - History - Print - Recent Changes - Search
Page last modified on July 14, 2015, at 03:53 PM