Identifying low risk of disease recurrence in localised ccRCC is key for gatekeeping in the adjuvant therapy enrolment. Uncertainty increases the number of patients required for accrual to achieve statistical power. Current scoring systems are good at identifying very low and very high risk cohorts, but have not been proved to be as effective at accurately predicting disease recurrence in intermediate groups. These patients are perhaps those likely to benefit from intervention in addition to surgery, but many may be treated unnecessarily. Using digital pathology, image analysis and machine learning we sought to stratify for risk in this intermediate category.
Definiens Tissue studio and Developer XD were utilised for object thresholding to measure tumour cell nuclear morphological features on H&E digitised images from 120 ccRCC training set from UK and 217 ccRCC validation set from Singapore. Multiplexed immunofluorescence (mIF) was performed on 120 cases to co-detect neo-vasculature and pan-T cells. An algorithm was derived to measure spatial relationships between blood vessels and T cells. A statistical model was developed by generalised linear model with spatially adaptive local smoothing algorithm, having specificity prefixed (0.8-1) plus cross validation.
Replacing manual nuclear grade with AI aided tumour cell nuclear morphological features improved the specificity of Leibovich score (LS) from 0.76 to 0.86 and from 0.84 to 0.94 in training and validation sets, respectively. Moreover, tumour microenvironment (TM) parameters significantly improved the specificity up to 0.93 in the training set. The negative predictive values of both LS 5 and 6 were zero, but by applying the algorithm the specificity for LS 5 and 6 cases became 0.93 and 0.40 respectively.
By applying image analysis it is possible to identify lower risk for recurrence patients in a conventionally identified intermediate risk group based on routine ccRCC H&E images, and multiplexed TM features. This approach to pathology should help refine selection of patients for clinical trials and form the basis of future AI-enabled prognostic and predictive algorithms in ccRCC.
Clinical trial identification
Legal entity responsible for the study
Renal Cancer Research Fund and NHS Lothian.
All authors have declared no conflicts of interest.