Automated classification of renal interstitium and tubules by local texture analysis and a neural network

OBJECTIVE: To segment renal interstitial space in order to automatically quantify renal cortical interstitial volume fraction (Vvint/cortex) by means of image analysis techniques. STUDY DESIGN: The study group consisted of 35 renal biopsies with different degrees of chronic interstitial damage. Biopsies were stained with Sirius red and digitized under polarized light. Two methods were employed to segment interstitial space: (1) interstitial bright particles were thresholded, and afterwards interstitial space was reconstructed with a morphologic operation, and (2) the texture of the surroundings of each pixel was quantified by means of local granulometry, and this information was employed as the input of a neural network in order to classify interstitial and tubular pixels. RESULTS: The correlation between Vvint/cortex obtained manually and both methods was r =.92. The first method produced some deformation of tubular contours and underestimated Vvint/cortex (β =.70) when compared to the second approach (β =.95) (P<.05). CONCLUSION: Two different algorithms based on image analysis techniques allow the classification of renal interstitial and tubular structures and consequently allow the automated and precise estimation of renal Vvint/cortex.