P1.09-11 Influences of Sampling Method to Morphological Feature Measurement of Lung Cancer Cell

Abstract

Progress of imaging technologies in the field of histopathology enables us to exploit artificial intelligence (AI) techniques to detect cancer based on digital images for screening or quality assurance of diagnosis process. Nowadays, reports on the application of AI to cancer detection which claim 99-percent detection accuracy are found in every proceedings or journal of digital pathology. However, little attention has been paid to the influences of sampling method to AI-based histological diagnosis. Whole slide images of hematoxylin and eosin (H&E) stained slides collected from 94 non-small cell lung cancer (NSCLC) cases were captured by a virtual slide scanner (NanoZoomer, Hamamatsu Photonics, Japan). Sampling methods were needle biopsy (59 cases), operation (12 cases) and endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA) (29 cases). Regions of interest (ROI) were selected by an experienced pathologist. After selecting tumor cells only by AI-based tumor cell detecter (Figure.1), following morphological features were calculated: nuclear area, perimeter (Peri), circularity (Circ) and five texture features, i.e., angular secondary moment (ASM), contrast (Cont), homogeneity (Hom) and entropy (Ent) of gray level co-occurrence matrix (GLCM), and contour complexity (CC). We found significant differences (p<0.05) in most of feature values except nuclear area and perimeter. Our results suggest that methods of sampling significantly affect morphological feature values of nucleus and this fact must be taken into consideration when applying AI-based techniques to tissue image classification.