How to Automatically Identify Regions of Interest in High-Resolution Images of Lung Biopsy for Interstitial Fibrosis Diagnosis

Abstract

Airway-centered Interstitial Fibrosis (ACIF) is a histological pattern of Interstitial lung diseases. Its diagnosis requires a multidisciplinary approach, in which diverse information, such as clinical data, computed tomography data, and lung biopsy data, is analyzed. Biopsy samples are digitized at high-resolution. Of crucial interest are broncho-and bronchiolocentric remodeling with extracellular matrix deposition. To analyze an image, specialists have to explore it at low microscope magnification, select a region of interest and export a smaller specified sub-image to be interpreted at higher magnification. This process is performed several times, requiring hours, becoming a tiresome task. We propose a method to support pathologists to identify specific patterns of ACIF in high-resolution images from lung biopsies. This can be done by a) automatic microscope magnification reduction; b) computing the probability of pixels belonging to high-density regions; c) extracting Local Binary Patterns (LBP) of the high-and low-density regions; and d) visualizing them in color. We have evaluated our method on nine high-resolution lung biopsies. We have tested the LBP features of high-and low-density regions with the kNN algorithm and obtained a classification accuracy of 94.4%, which is the highest one reported in the literature for this type of data.