Textural analysis of multiparametric MRI detects transition zone prostate cancer
Harbir Singh Sidhu1, Salvatore Benigno1, Balaji Ganeshan2, Nikos Dikaios1, Edward William Johnston1, Clare Allen3, Alex Kirkham3, Ashley M Groves2, Hashim Uddin Ahmed4, Mark Emberton4, Stuart A Taylor1, Steve Halligan1, and Shonit Punwani1

1Centre for Medical Imaging, University College London, London, United Kingdom, 2Institute of Nuclear Medicine, University College London, London, United Kingdom, 3Radiology, University College London Hospital, London, United Kingdom, 4Research Department of Urology, University College London, London, United Kingdom


Transition zone (TZ) prostatic tumors are more difficult for radiologists to detect on multiparametric MRI compared with peripheral zone tumors and are systematically undersampled by conventional transrectal ultrasound biopsy.

Assessment of whole TZ heterogeneity by spatially filtered textural analysis of routinely acquired multiparametric MRI images can discriminate significant tumor without the need to predefine tumors and at no additional burden to patients. TZ containing significant tumors show reduced entropy on coarsely filtered early post contrast T1 and T2 weighted images and reduced kurtosis unfiltered ADC values.

In the future, this could augment radiological interpretation and facilitate computer-aided diagnosis.


To evaluate multiparametric MRI (mpMRI) derived selective-scale filtration-histogram textural analysis parameters from the whole transition zone (TZ) for diagnosis of clinically significant TZ prostatic tumor.


TZ tumors remain difficult to appreciate on mpMRI studies [1], with reported sensitivity/specificity for detection of 0.53/0.83 compared with 0.80/0.97 respectively for peripheral zone (PZ) tumors [2].

Tissue heterogeneity data afforded by MR textural analysis (MRTA) could increase the diagnostic accuracy of radiologists in detecting TZ tumors with potential utility in computer aided diagnosis (CAD).


Institutional review board waived the informed consent requirement for this retrospective study.

Men with suspected prostate cancer undergoing mpMRI prior to ‘20 zone’ template prostate mapping (TPM; [3]) biopsies between 1st January 2010 to 31st December 2012 inclusive were identified (n=210). Those with peripheral zone tumor, biopsy within 6 months prior to mpMRI, previous treatment for prostate cancer, metallic hip prostheses, or incomplete mpMRI/TPM data sets were excluded (n=143).

Sixty-seven men (mean age 63.4 years) underwent 1.5T mpMRI (Fig 1) prior to TPM (median interval 56 days). Twenty-six men (39%) had significant TZ tumor.

Subjects were grouped according to a previously used definition of cancer significance [4] categorising clinically significant disease as ≥ Gleason 3+4 OR ≥4mm maximum cancer core length (MCCL); other disease (i.e. ‘low risk’ <4mm MCCL AND ≤Gleason 3+3) was classified as non-significant and grouped with benign TZ histology [5].

Two experienced radiologists in consensus matched TPM to the single axial slice best depicting tumor or the largest TZ AP diameter for those with benign histology. A third blinded radiologist contoured the entire TZ as a region of interest (ROI) on each matched ADC, T2 and early post-contrast T1 weighted slice.

Whole TZ ROIs underwent MRTA using proprietary 'TexRAD' research software (version 3.3, TexRAD Ltd, Feedback Plc, Cambridge UK) [6] first using a Laplacian of Gaussian (LoG) band-pass spatial scale filter (SSF) to highlight features ranging between 2mm (fine) and 6mm (coarse) in radius (Figure 2) to extract and enhance imaging features at different sizes and highlight intensity variation corresponding to fine, medium and coarse texture scales within the TZ. Histogram analysis quantified first-order statistics of entropy, skewness, kurtosis of the TZ prior to and following filtration.

Differences between TZ ROI textural parameters were analysed using Mann Whitney U test (statistical significance p<0.05) and diagnostic accuracy of parameters for detection of significant TZ tumor determined by receiver operating characteristic (ROC) area under curve (AUC) analysis. The combination of the most discriminatory parameters for individual mpMRI sequences was determined using multivariate ROC-AUC analysis [7].


All analyses were performed on a whole TZ basis (incorporating tumor where present). Figure 3 shows the median values, interquartile ranges and significance of differences for individual best performing textural parameters prior to/following application of incremental spatial filters for each MRI sequence.

TZ containing significant tumor compared with benign/non-significant TZ demonstrates significantly higher early post contrast T1 and T2 image homogeneity particularly when images undergo prior ‘coarse’ spatial filtration (6mm), with entropy values of 5.90 and 5.69 respectively versus 6.16 and 5.94 (p<0.001).

Analysis of unfiltered TZ ADC histogram values in TZ containing significant tumor reveals significantly less peaked distribution compared to benign/non-significant TZ, with ADC kurtosis value of -0.51 versus 0.09 (p<0.001) becoming less significant with any prior spatial filtration (Figure 4). Median ADC, T1 and T2 skewness did not demonstrate any consistent difference between the two groups.

Coarsely filtered early post contrast T1 and T2 weighted image entropy and unfiltered ADC kurtosis are the best classifiers of TZ containing significant tumor for each respective multiparametric MR (mpMRI) sequence with area under the curve of 0.82, 0.77 and 0.80 respectively and in combination 0.89 (Figure 5).

Discussion and Conclusions

Classification of TZ tumor-containing slices by textural analysis (ROC-AUC 0.77 to 0.82) was comparable with previously reported visual detection of TZ tumor by radiologists (ROC-AUC 0.73 to 0.84) [8] and confirms utility of increased TZ homogeneity on T2 and post contrast T1 images as a good discriminator of cancer [9] and suggests less peaked ADC values in TZ containing significant tumor. Applying spatial filters to images prior to textural quantification further improves classification performance on higher resolution T2/T1 images though not on lower resolution ADC images.

To our knowledge, this study is distinct from other work as it assessed in-vivo textural features derived from each mpMRI sequence from the entire TZ which obviates inherent difficulties in small lesion pre-identification and contouring.

Assessment of TZ heterogeneity by spatially filtered textural analysis of routinely acquired mpMRI images could therefore augment radiological interpretation and may in the future facilitate CAD.


This work has been supported by the KCL-UCL Comprehensive Cancer Imaging Centre funding [Cancer Research UK (CR-UK) & Engineering and Physical Sciences Research Council (EPSRC)].The majority of this work was undertaken at University College London Hospital and University College London, which receive a proportion of funding from the NIHR Biomedical Research Centre funding scheme [Department of Health UK].


1. Langer DL, van der Kwast TH, Evans AJ, et al. Prostate cancer detection with multiparametric MRI: logistic regression analysis of quantitative T2, diffusion-weighted imaging, and dynamic contrast-enhanced MRI. J Magn Reson Imaging 2009; 30:327–33.

2. Delongchamps NB, Rouanne M, Flam T, et al. Multiparametric magnetic resonance imaging for the detection and localization of prostate cancer: combination of T2-weighted, dynamic contrast-enhanced and diffusion-weighted imaging. BJU Int. 2011; 107(9):1411-811.

3. Onik G, Barzell W. Transperineal 3D mapping biopsy of the prostate: an essential tool in selecting patients for focal prostate cancer therapy. Urol Oncol 2008; 26:506–510.

4. Lecornet E, Ahmed HU, Hu Y, et al. The accuracy of different biopsy strategies for the detection of clinically important prostate cancer: a computer simulation. J Urol. 2012; 188(3):974-80.

5. Stamey TA, Freiha FS, McNeal JE et al: Localized prostate cancer. Relationship of tumor volume to clinical significance for treatment of prostate cancer. Cancer 1993; 71:933-938.

6. Miles KA, Ganeshan B, Hayball MP. CT texture analysis using the filtration-histogram method: what do the measurements mean? Cancer Imaging 2013 Sep 23; 13(3):400-6.

7. Hanley JA, McNeil BJ. The meaning and use of the area under a Receiver Operating Characteristic (ROC) curve. Radiology 1982; 143:29-36.

8. Oto A, Kayhan A, Jiang Y, et al. Prostate cancer: differentiation of central gland cancer from benign prostatic hyperplasia by using diffusion-weighted and dynamic contrast-enhanced MR imaging. Radiology 2010; 257(3): 715–723.

9. Akin O, Sala E, Moskowitz CS, et al. Transition zone prostate cancers: features, detection, localization, and staging at endorectal MR imaging. Radiology 2006; 239(3): 784–792.


mpMRI sequence parameters; *dynamic contrast enhanced MRI – 0.2 ml/kg intravenous gadolinium contrast agent injected at 3 ml/s followed by 20ml saline flush; T2w TSE – T2 weighted turbo spin echo; EPI-DWI – echo planar imaging - diffusion weighted imaging; FLASH – fast low angle shot.

Axial T2w image with contoured transition zone showing unfiltered image in top left pane with increasingly ‘coarse’ filtering (spatial scaling factors- 2mm top right, 4mm bottom left and 6mm bottom right panes) in patient with anterior transition zone tumor (Gleason 3+4; maximum cancer core length 5mm) prior to histogram analysis.

Median values ( interquartile range 25%-75%) for unfiltered/filtered (spatial scaling factor; SSF) early post-contrast T1 and T2 weighted entropy and ADC map kurtosis derived from first order histogram analysis. P-values have been calculated using Mann Whitney U test. Bold indicates most significant difference in values (i.e. non-significant/benign vs significant tumor TZ).

Box plots showing best performing discriminators for each MR sequence using T2 entropy (filter 6; spatial scaling factor 6mm), post-contrast T1 (filter 6) and ADC kurtosis (unfiltered; no spatial scaling factor). Box indicates interquartile range; line indicates median and whiskers most deviated data points. Mann Whitney U p-values also shown.

Receiver operating characteristic (ROC) curves of the different textural features for discrimination of significant prostatic tumor containing transition zone ROIs from non-significant TZ ROIs with area under curve (AUC) values as shown (‘fltr' refers to spatial scaling factor filter application).

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)