Classification of Lung Cancer Using Nuclear Chromatin Characteristics

Abstract

Background/Objectives: Studies on automatic diagnosis to assist pathologists have been accelerating due to artificial intelligence technology. We investigated whether two types of lung cancers, squamous cell carcinoma (Sca) and neuroendocrine tumor (NET), can be automatically classified with only the features of nuclear chromatin, similar to the method for diagnosing cancer by a pathologist.

Methods/Statistical analysis: In order to classify lung cancer with only nuclear chromatin features, it is necessary to extract texture features of the nuclear chromatin area or train a deep learning model with only the nuclear chromatin area as input. We developed a segmentation algorithm that automatically extracts nuclear areas based on the k-means algorithm to automatically generate training and testing datasets for deep learning models. And using four types of deep learning models, we performed lung cancer classification using only nuclear chromatin features. 

Findings: The pathologist judges the presence and type of cancer by looking at the size and placement of the nuclei in the cell, and the texture characteristics. In this paper, a data set for a deep learning model was created by extracting only the nuclear region from a pathological image with a 40x magnification of WSI using an automatic segmentation algorithm. For four classification models, an average of 84% classification accuracy was obtained. Since only the nuclear region was used for training and testing, the deep learning model extracted and classified the size, shape, and texture features of the nuclei. Therefore, the classification of lung cancer performed in this study is similar to the method that the pathologist classifies cancer based on the features of nuclear chromatin. Since it showed 84% classification accuracy for the four types of models, it was proved that the automatic classification system using deep learning technology can be used to assist pathologists' diagnosis.

Improvements/Applications: The accuracy of a cancer diagnosis system that applies only deep learning models has limitations. To improve diagnostic accuracy, we are going to study on a model that combines a deep learning model and machine learning technology that extracts and uses biomarker features for analyzing nuclei texture patterns.