By: Dan Sperling, MD

The name of Robert Haralick (b. 1943) may not be familiar, but in the world of computer imaging and image analysis, his contributions to computer vision have practical applications in many fields including biology and medicine. As computer science and digitized imaging evolved, Haralick was interested in analyzing small black-and-white image areas using texture and tone, based on the varying shades of gray in the image.  By assigning gray-tone values to each resolution cell, image-processing tasks can be performed by a computer, e.g. coding, restoration, enhancement and classification. At age 30, after working with NASA, Haralick published a paper dealing with “the task of developing a set of features for classifying or categorizing pictorial data” such as identifying an agricultural crop category from satellite photos of the earth’s surface.[i] In particular, his technique of texture analysis improved information about observed surfaces through features characterized as smooth, regular, rippled, lineated, etc. To illustrate this principle, Haralick noted that “…in the humid tropics, fine texture on radar imagery can be indicative of nonresistant fine-grain sedimentary rocks…while a coarse texture can be indicative of coarser grained sedimentary rocks.”

Detecting cancerous tumors through imaging has gained much ground as a precursor to biopsy. While diagnostic imaging such as x-rays, mammograms, CT or PET/CT scans always involve some ionizing radiation, no matter how minimal, that damages DNA molecules in healthy tissue, magnetic resonance imaging (MRI) produces no ionizing radiation at all. Since MRI can discriminate tissue differences, medical science continues to explore ways to its ability to detect cancer. Research integrating Haralick texture analysis with multiparametric MRI (mpMRI) is breaking new ground.

For instance, one study on detecting very small breast tumors successfully incorporated Haralick textural features into dynamic contrast enhanced MRI (DCE-MRI).[ii]  The authors applied texture analysis to five time-sequenced gray scale images after administering a contrast agent (to record its uptake and washout by cancer cells) in an approach called dynamic texture quantification. This resulted in better ability to classify what they called “diagnostically challenging lesions.” While acknowledging certain limitations, this 2012 study was a step closer to computer-automated classification of tissue differences.

With regard to prostate cancer (PCa), the incorporation of Haralick textural features is highly promising in distinguishing both significant and insignificant disease in anatomic prostate zones that are reported to “hide” PCa from mpMRI.[iii] A new study out of Memorial Sloan Kettering Cancer Center (New York, NY) was designed to investigate Haralick texture analysis to differentiate Gleason scores in suspicious lesions identified on imaging.[iv] The study involved 147 patients who had both T2-weighted MRI (T2WI) and diffusion weighted MRI (DWI-MRI; for an explanation of DWI as well as DCE, see https://sperlingprostatecenter.com/mri-imaging-prostate-cancer-two-parameters/). In processing the resulting images, the following texture features were extracted and analyzed: Energy, Entropy, Correlation, Homogeneity, and Inertia. Radiology readers used T2WI and DWI-MRI to identify cancers ?0.5 ml as well as non-cancerous peripheral zone (PZ) and transition zone (TZ) tissue. After the patients underwent radical prostatectomy, the prostate specimens were used as the comparison reference with the images.

The researchers identified 143 PZ cancers and 43 TZ cancers. In terms of the Haralick texture features, the authors report that the imaging texture of PZ tumors showed higher Entropy and Inertia with lower Energy, Correlation and Homogeneity compared to non-cancerous tissue. In the 43 TZ cancers they identified, DWI-MRI illuminated significant differences for all five texture features, and T2WI revealed significant differences for Correlation and Inertia. When compared with the findings from the surgically removed specimens, Gleason score (GS) “was associated with higher Entropy…and lower Energy.” When DWI-MRI images were analyzed and compared with gland specimens, Energy and Entropy were “significantly different in GS ?3?+?4 versus ?4?+?3 cancers…” The authors concluded, “Several Haralick-based texture features appear useful for prostate cancer detection and GS assessment.”

Haralick’s work with satellite images of the earth’s surface is bringing down-to-earth imaging results that bring mpMRI ever closer to becoming a PCa diagnostic modality.

[i] Haralick R, Shanmugam K, Dinstein D. Textural features for image classification. IEEE Transactions on Systems, Man and Cybernetics. 1973 Nov;SMC-3(6):610-21.

[ii] Nagarajan M, Huber M, Schossbauer T et al. Classification of small breast lesions in MRI: evaluating  the role of dynamically extracted texture features through feature selection. J Med Biol Eng 2012;33(1):59-68.

[iii] Rosenkrantz A, Verma S, Turkbey B. Prostate cancer: tope places where tumors hide on multiparametric MRI. Am J Roent 2015 April;204(4). Online exclusive.

[iv] Wibmer A, Hricak H, Gondo T et al. Haralick texture analysis of prostate MRI: utility for differentiating non-cancerous prostate from prostate cancer and differentiating prostate cancers with different Gleason scores. Eur Radiol. 2015 May 21. [Epub ahead of print]