Computational image analysis in clear cell renal cell carcinoma

Abstract

Background: In the UK, kidney cancer is the most lethal urologic cancer whose major subtype is renal cell carcinoma (RCC). Surgical resection is the first line of the treatment for renal cell carcinoma (RCC). A number of different integrated staging systems such as UISS, SSIGN and Leibovich score, has been introduced and utilised in clinic as a prognostic tool or as an inclusion criterion of clinical trials. Among them, Leibovich score has been widely utilised in the UK to predict the likelihood of disease free survival for clear cell renal cell carcinoma (ccRCC). However, its prediction rate of disease relapse free for 5 years after surgery varies from 97% (low risk) to 31% (high risk) which might lead to inclusion of some ccRCC patients, who would not recur, into a clinical trial. Therefore, we aim to improve the prediction power of recurrence free (specificity) in localised clear cell renal cell carcinoma (ccRCC) patients, by either modifying Leibovich score with more precise and accurate measurement of ccRCC nuclear morphological features or by improving currently available Leibovich score with the features measured from the chromatin marker colocalisation status, chromatin marker Haralick texture features, and the tumour microenvironment using computational image analysis.

Methods: To modify Leibovich score by replacing manual Fuhrman’s nuclear grade with the computational image analysis measurement of the nuclear morphological features, digitised images from haematoxylin and eosin stained slides were utilised. For the chromatin marker (H3K9me3, H3K4me3 and HP1α) colocalisation analysis, Haralick texture analysis and tumour microenvironment marker (CD105 and CD3) analysis, the multiplexed immunofluorescence (IF) was performed and IF images were utilised. The image analysis was performed by using Definiens Tissue studio® (Definiens AG, Munich, Germany) and the Developer platform. Moreover, a novel statistical model was developed using AUCp (partial area under the curve) function in R studio, which defines the range of specificity between 1 and 0.8 on the basis of the binomial GLM (Generalised linear model) framework with GAM-SALSA (Generalised Additive Models - Spatially adaptive local smoothing algorithm) as a calibration tool. In order to avoid the overfitting problem, 5-fold cross validation with 100 times repetition was also added in the analysis.

Results: Firstly, our statistical model replaced Fuhrman’s nuclear grade with ‘mean perimeter’, which was named ‘Modified Leibovich algorithm’. The modified Leibovich algorithm improved its overall specificity 0.86 (80 out of 93 cases) from 0.76 (71 out of 93) from the classic L score in the Scottish training cohort. In particular, the most increase in specificity was seen in Leibovich score 5 and 6, which were 57% and 40%, respectively. The modified Leibovich algorithm also increased overall specificity up to 0.94 (141 out of 150 cases), compared to the classic (original) Leibovich score whose specificity was 0.84 (126 out of 150 cases) in a Singaporean validation cohort. Moreover, specificity was dramatically increased in Leibovich score 5 from 0% to 92%. Secondly, the chromatin marker colocalisation feature significantly improved the specificity of the classic and partial Leibovich score up to 0.98 in overall. In particular, the specificity of the cases in Leibovich score 4, 5 and 6 was improved up to 100%. However, the chromatin marker Haralick texture features did not improve the specificity of the classic and partial Leibovich score as much as the chromatin marker colocalisation features. Thirdly, the tumour microenvironment features such as the density and the spatial distances of CD105 positive blood vessels and CD3 positive mature T lymphocytes augmented the specificity of the classic and partial Leibovich score up to 0.93. In particular, it improved the specificity up to 92% in Leibovich score 5 compared to the classic Leibovich score.

Conclusions: Computational image analysis enabled to measure such various ranges of features not only in tumour cells, but also in tumour microenvironment. In this study, ccRCC tumour cell nuclear morphological features, chromatin marker colocalisation feature, chromatin Haralick texture features and tumour microenvironment features were visualised by the multiplexed immunofluorescence and measured by the Tissue studio and developer software: this cannot be done manually. The data from this analysis significantly augmented the specificity of the currently available prognostic tool, Leibovich score, in clinic. In particular, the very last final model developed by combining features of the chromatin marker colocalisation (‘Average.HP1a.Intensity..H3K4_HP1a.Overlap.’ and ‘Manders.Coefficient.Nuclei..H3K4.HP1a..M2’) and tumour microenvironment (‘Mean.CD3.Area.percentage.of.All.Tissue’ and ‘Mean.No..of.CD105’) selected along with the classic Leibovich score, predicted 100% (93 out of 93 cases) correctly the cases which did not experience the disease recurrence within 5 years after surgery, while the classic Leibovich score predicted 76% (71 out of 93 cases) correctly. This could have prevented 22 ccRCC patients not only to get unpleasant and unnecessary treatment after nephrectomy, but also helped them having been free from the fear of disease recurrence.