A Novel CAD System for Detecting Acute Rejection of Renal Allografts Based on Integrating Imaging-markers and Laboratory Biomarkers

The goal of this paper is to determine the parameters that are correlated with the biopsy diagnosis of acute renal rejection (AR) post-transplantation, using laboratory biomarkers and (3D+ -value) diffusion weighted MR (DW-MR) image- markers. 16 patients with non-rejection (NR) and 45 patients with AR renal allografts determined by their renal biopsy as a gold standard were included. All kidneys were evaluated using both laboratory biomarkers (e.g., creatinine clearance (CrCl) and serum creatinine (SCr)) and DW-MR image-markers. To extract the latter, DW-MR kidney images were first segmented using a geometric deformable model, then, DW-MR image-markers known as apparent diffusion coefficients (ADCs), were estimated for segmented kidneys at multiple b-values (i.e. strength and timing, of, the, field, gradients, (b50,b 100,,…,b 1000 s/mm2)). A statistical analysis investigating possible correlations between potential biomarkers of AR and the biopsy diagnosis was firstly performed. Two categories of parameters were mainly examined: (i) laboratory biomarkers (CrCl and SCr) and (ii) the average ADC (aADC) at individual b-values. Analysis of Variance(ANOVA) and the likelihood ratio ($\chi^{2}$) tests found that both CrCl and SCr affected significantly the likelihood of AR, as did the aADC for the individual b-values of b 100, b 500, b 600, b 700, and b 900s/ mm2. Nevertheless, patient demographics (i.e. age and sex) and the aADC at the remaining b-values had no significant effect. The statistical analysis results encouraged us to investigate if this can lead to building a computer-aided diagnostic (CAD) system with the ability to classify AR from NR renal allografts. To achieve this goal, stacked auto-encoders (SAEs) based on deep learning approach were trained using the fusion of the statistically significant DW-MR image-markers and laboratory biomarkers for the classification purposes. Preliminary results obtained (92% accuracy, 92% sensitivity, and 94% specificity) hold a lot of promise of the presented technique to be reliably used as a noninvasive post-transplantation diagnostic tool.