“Textural analysis of multiparametric MRI detects transition zone prostate cancer”

This study evaluated the diagnostic accuracy of textural parameters, derived from clinical prostate mpMRI, for detection of TZ cancer. Previous work has confirmed that quantitative mpMRI parameters (e.g., ADC) can differ between benign and cancerous TZ regions [18]. Unlike previous studies, here we derived single-slice whole TZ textural parameters (including cancer pixels where cancer was present) and evaluated differences in the histographic pixel distribution between patients with and without significant cancer.

We found that textural features from an image of the entire TZ are altered significantly, when containing even a small proportion of significant cancer, which no longer holds true (for best performing textural parameter) when the same tumour is excluded from the analysis. Overall, classification of TZ tumour-containing slices by best performing single textural parameter and/or bivariate combination (ROC-AUC 0.80 to 0.86) was comparable with previously reported visual detection of TZ tumour by radiologists (ROC-AUC 0.73 to 0.84) [19, 20].

Kurtosis is a measure of histogram “peakedness”. Positive kurtosis indicates a more peaked distribution of pixel signal intensities. We found reduced ADC kurtosis was the best univariate classifying textural feature (ROC-AUC 0.80 on ADC images) on a whole TZ basis and demonstrated a higher ROC-AUC than the non-textural parameter of ADC mean (ROC-AUC 0.71 on ADC images). A larger cohort is required to test the statistical significance of this difference. We expect textural measures, based on the relationship of pixels in a given ROI rather than absolute pixel intensities, are likely more robust to variations between individuals and scanners. Whilst kurtosis measures from T2 and early post-contrast T1 weighted images performed less well for the detection of cancer. In spite of the relatively modest ratio of significant tumour compared to remaining TZ area (~5 %), we have found that exclusion of this radiologically visible tumour from the single-slice whole TZ ROI results in loss of significance in the difference in ADC kurtosis between the two groups. We posit this to be the result of elimination of a second differing signal intensity ‘population’ (i.e., a low ADC significant tumour) from the TZ ROI resulting in normalisation of kurtosis when compared to the non-significant/benign cohort reflected in more ‘peaked’ values. Wibner et al. [21] have previously demonstrated the utility of ADC-derived textural parameters in TZ tumour detection and differentiating grade based on lesion analysis. The textural analysis approach may obviate previously described efforts to normalise by using ADC ratios (normalized to non-tumorous tissues) [22], perform whole lesion measurement and/or generate histogram analysis including median and low percentile ADC values [23].

Entropy is a measure of image ‘busyness/irregularity’ and, as such, the observation of significantly lower mean TZ entropy values in patients with significant cancer would reflect increased overall signal homogeneity. Post-contrast T1 weighted image entropy was the second best univariate classifier, in keeping with previously observed homogenous enhancement features of TZ tumours [20]. However, the observation that the significant difference in T1 entropy between TZ containing significant tumours and non-significant/benign TZ persists even after the exclusion of radiologically visible tumours in the former group reflected more homogeneous enhancement throughout the whole TZ in these cases. The mechanism for this is unclear, though growing evidence implicates chronic inflammation as a contributor to prostate cancer development and progression [24] and present in benign prostate tissue associated with high-grade prostate cancer [25]. The finding of more homogenous enhancement in the ‘non-malignant’ parts of TZ harbouring significant tumours may reflect these aspects of prostate cancer ontogeny. ADC entropy was also significantly lower in single-slice TZ ROIs containing tumours and this relationship persisted following removal of cancer from the TZ ROI possibly due to the same/similar mechanism. Additionally, at the time these data were acquired, as demonstrated in Table 2, the lowest b-value used for calculation of the ADC map was 0 sec/mm2, and it is possible the findings of ADC entropy may at least in part reflect a perfusion difference component. These findings may not have been appreciated in prior quantitative imaging studies, most of which concentrate on assessing the vascular properties of tumours themselves, and further confirmatory (histologically correlative) studies are required.

Several T2 weighted features aid visual identification of TZ tumours; including homogeneously reduced lesion signal intensity, ill-defined lesion margins and lenticular lesion shape [19]. However, on a whole TZ basis, the current study does not confirm that quantitatively reduced whole TZ entropy (increased homogeneity) on T2 weighted images is a particularly good discriminator. This may relate to the relatively small area of tumours compared with background non-tumorous TZ and/or relate to the study size. Additionally, ADC and early post-contrast T1 images reflect processes occurring on a ‘microscopic’ level and can reflect field changes (for example, subtle inflammatory change) that may explain the significance of differences between benign TZ and that harbouring significant tumours. Whilst several studies have described homogenous reduced tumoral T2 signal compared to benign TZ, the remaining (and predominant approximately 95 % by area) ‘normal’ TZ glandular tissue is variably heterogeneous across patients and images are more reflective of macroscopic differences and subtle changes in T2 signal induced by background inflammation may be masked.

Skewness is a measure of histogram asymmetry; a zero value indicates a symmetrical distribution around the mean. Skewness was a relatively poor classifier of TZ cancer when applied to any image sequence. It has been speculated [23] that in prostate tumours densely packed with malignant cells, the resulting histograms from tumour ROIs are likely to be less skewed compared with tumours with more heterogeneous cellular density. However, we evaluated skewness for all pixels within the whole TZ, and the proportion of tumour pixels may have been insufficient to influence skewness significantly particularly given such variation in cellular density is not captured by traditional histological grading [26].

To our knowledge, this study is distinct from other work as it assessed in vivo textural features derived from each mpMRI sequence for the entire TZ. The study was conducted as a proof of concept for the purpose of detecting significant TZ tumours. Future work will examine the translation of this concept over whole TZ volume to examine utility in augmenting visual radiologist assessment (e.g., added value of prompting readers to areas for examination which may contain significant cancer) and/or automated detection. Interpretation of images by radiologists is complex [27] and, although an assessment of texture is made, to date this visual textural assessment has not been characterised objectively. We believe that textural analysis of each multiparametric image is analogous to the manner by which radiologists visually localise TZ cancer on mpMRI [4]. Other workers [e.g., 28, 29] have examined first and second order (e.g., two-dimensional grey-level co-occurrence matrix) textural features from one or more mpMRI sequences from ROIs drawn around individual TZ tumours. Such approaches, while informative, have less clinical application, since they are more computationally intensive and require the radiologist to first identify areas of concern. Furthermore, where textural features of small lesions are evaluated at larger spatial scales, these features can be biased significantly by region boundary. In contrast, assessment of the entire TZ texture requires less development as a clinical diagnostic tool (i.e., only segmentation of the TZ from the PZ, which may be an automated/semi-automated procedure [30]) and minimizes boundary-related effects.

Our study has some limitations. All mpMRI images were acquired on a 1.5 T scanner and generalizability to 3 T platforms was not assessed. However, our 1.5 T mpMRI protocol was in keeping with recommendations from a European Consensus meeting [31]. We did not use endorectal coils, and note that the European Consensus Meeting failed to agree regarding their necessity even at 1.5 T. We routinely perform DCE MRI using a higher spatial resolution and lower temporal resolution than some other centres as advocated by recent guidelines [3, 32]. Therefore, we expect the generalizability of post-contrast T1 weighted textural feature findings to be limited to centres that similarly maintain higher spatial resolution. We did not have access to radical prostatectomy specimens as a reference ground-truth. However, employment of a TPM based reference standard [33], which is a significant improvement over conventional TRUS biopsy [14], can help avoid the spectrum bias towards more severe cases that occurs with studies when prostatectomy specimens are mandated. We acknowledge that there could be some error when registering TPM histopathology to mpMRI. However, it should be noted that prostatectomy specimens themselves are not free from registration errors induced by processing (shrinkage, distortion, and tissue-loss) [34]. Lastly, we manually contoured the single-slice TZ ROIs though we do not expect this to unduly affect the results and envisage automated segmentation could be used as available.

Our results have immediate clinical relevance, and confirm textural features may inform CAD software to highlight images on which significant tumours could be present within the TZ. In developing this whole TZ approach using MRTA as an imaging biomarker, it is recognised that a series of studies would be required to further validate the findings as per various ‘roadmap’ approaches in contemporaneous guidelines (http://​www.​cancerresearchuk​.​org/​sites/​default/​files/​imaging_​biomarker_​roadmap_​for_​cancer_​studies.​pdf). This approach could be adapted to several scenarios, subject to demonstration of efficacy, for example, by drawing attention to slice(s), which may require further examination for radiologists, probability of significant tumours being present across entire TZ and predictive information if there are relevant ‘field’ changes across the TZ. The United Kingdom NICE guidelines have adopted mpMRI for detection of prostate cancer in patients with a negative non-targeted TRUS biopsy (http://​www.​nice.​org.​uk/​guidance/​cg175/​chapter/​recommendations). Our textural evaluation technique may have particular relevance for such patients who are more likely to have TZ tumours that are systematically undersampled by TRUS [34].