I would make four points:

1. Texture in images can be invisible to the human eye. Computer texture analysis "sees" images quantitatively, and can discriminate low-contrast texture that is below the contrast sensitivity of the human eye, and it can tolerate much more extreme contrast (greater dynamic range of image intensity) than monochromatic image rendering of an image for visual inspection generally allows.

2. Texture generally depends on the length-scale of analysis. Natural patters are generally "fractal" --- i.e., what you measure quantitatively generally depends on the length-scale at which it is measured. (See Mandelbrot, "Fractal Geometry of Nature").

3. Texture analysis can "see" statistical properties of pixel-to-pixel variation that may be invisible to the eye. Statistical properties in images depends in part on the length-scale of analysis (sliding window size).

4. Much depends on the processing applied transforming the raw sensor (raw image data) into its visual rendering for human inspection. Generally this is a non-linear transformation. It may compress the dynamic range of the raw sensor data in some respects, and it may contrast-enhance  "interesting" image texture for visual inspection in other respects.  You do not tell us what sensor you are using, whether x-rays, ultrasonics, infrared, etc. -- so I can only speculate here. Note that there is nothing fundamentally visual about non-optical sensors that function outside of the visual band. For instance, the output of ultrasonic imaging is in the first place ultrasonic intensity, which is not a visual optical quantity.  At some stage, the output of the sensor must therefore be transformed into monochromatic or color rendering suitable for presentation in the human eye, on computer screen or in print, using what is ultimately an arbitrary non-linear transformation, transforming raw sensor output into visual color/contrast scale. This transformation can enhance or suppress interesting image features for human vision. If the transformation is unique and continuous (the mapping from raw data to image is "one-to-one and onto"), then all of the texture information is preserved, and it can be revealed by computer texture analysis, but it may be invisible to the eye. Ideally your imaging sensor provides raw data, and you can change the non-linear transformation yourself, to make some invisible textures easily visible.

I hope that helps.

Ronald

Andrés,

I guess Ronald has addressed your question in a more elaborative way. I see that your post embeds within two distinct parts; the second part of the question concerning the magnification effect on textural analysis is beyond my affirmative knowledge though I have some speculation. I appreciate that you wanted to make your concern as clear as possible but you have not revealed what sensor and texture metric you are referring to. I have, however, a wee comment to add on top of what Ronald has said regarding the first part. When you say: “two apparently similar parts” and “texture features find differences among the same visually homogeneous texture”, how are you quantifying the similarities between those patches/objects? I can only think that you are relying on visual similarities by eyeballing the two objects. It is worth reiterating what Robert mentioned that applying non-linear (or linear) transformations can reveal some structures to the human eye which where otherwise invisible especially when the tumour is projected as an oversaturated patch. Such transformations encompass the well-known contrast limited adaptive histogram equalization, Gabor, Hessian, entropy, range, and standard deviation filtering to name a few. Since textural metrics are, by design, quantitatively sensitive to statistical texture changes, I bet that you will always end up having differences in texture and the job of the image analyst is to capitalise on other side information to yield a decision whether the found differences are statistically significant and thereby be able to classify tumours (in your case).

Thank you for all your answers. I work with MRI mainly and use texture metrics derived from the co-occurrence matrix, gabor filters, LBP among others. My concern was because we usually try to find some differences in order to discriminate between different types of tumors but doing this we asume that all tumors corresponding to one class has the same/similar textures, but if I dig in comparing texture features among the asumed same class I do find features that show statistical differences.

Whether you are using the co-occurrence matrix, Gabor filters or the local binary pattern (LBP = LPB?), the classification is affected by the features extraction and selection in your final statistical model. What could improve the robustness of intra-class features is to ensure that images are normalised before the features are extracted to eliminate systematic differences between images. Another thing which you have not talked about, and which could hinder arriving at meaningful and robust features, is the way tumour regions are localised. Check if there is any discrepancy in this regard; say the automatic segmentation allowing for surrounding tissue interference. Recalling the magnification issue which you raised that it could be a problem, why not? That is the very reason why I emphasised on the standardisation of the images as a pre-processing step. This is what I have in mind at the moment, and I hope it is of some help. Hopefully you can hear from others.

Ola Andrés,

Very interesting that you find different textures in tumors that are classified the same. What has pathology to say about it?

Sure, textures depend on scale. A dress with a pied-de-poule pattern just looks spotty at an intermediate distance, and homogeneous at a large distance. So, doing texture analysis at several resolutions makes sense. I would not magnify the original data though: textures you then find could (partially) be introduced by the interpolation algorithm you used. Do the analysis in millimeters instead of pixels; then there is no need to standardize image resolution as suggested above.

While it is true that texture analysis can reveal differences that were unnoticed, I have experienced the other way around too: given the right level and window setting, the 'formidable' human eye sees differences that are very hard to incorporate in a computer program.

Hi Andrés, I have used the Gabor Filter Algorithm in order to identify texture of eyelashes in bio images, it has being usefull, Let me find some papers about it to share with you.

Intra-class texture's “slight” differences should be tolerated by the classifier. It is not uncommon for oncologists/pathologists to have discrepancies in their reading/interpretation/assessment of a given sample. At least in my field of mammography where the BIRADS, a quality control system which is designed to standardize reporting, is extensively used. Even among experienced radiologists, the BIRADS visual assessment of breast composition (categorisation of mammograms into different density groups), for example, may differ as reported in a recent research paper.

Andres,

Texture information contain in an image will very much be depended on the spatial resolution of the image and the type of feature being imaged.  For example, in the remote sensing community (which is where I come from) we work with satellite images that have 1km resolution/pixels and some images with 0.5m resolution.  They both contain texture information that is related to the resolution of the pixels.  Texture typically relates to the high-frequencies in an image vs the more broad low-frequencies.  In my case I have investigated ways to extract the texture information of such images (which can be optical, microwave/radar, and acoustic/sonar of under the ocean) but use the same type of methods knowing that it is the high-frequency components of the image that I am interested in.  In my case I used High Pass Filters (HPFs) to extract the spatial variability at local (3x3 to 15x15), intermediate (21x21 to 51x51), and regional (75x75 to 501x501) pixel windows about the center pixel.  The extracted 3 Spatial Variability Images (SVI) can then be used as input to classification techniques.

Examples of this are given in the papers linked to below (within Research Gate):

* An automatic Optimum kernel-size selection technique for edge enhancement

* Comparison of Spatial Variability in Visible and Near-Infrared Spectral Images

 * Extraction of spatial information from remotely sensed image data - An Example: 

  GLORIA sonar images

Pat