Empirical study on the assessment of deep learning based lung cancer histopathology image classification

Histopathology images of the lung biopsy tissues is considered as the gold standard modality to confirm the presence and type of cancer. However, owing to the complex morphology, texture and presence of artifacts during sample preparation, staining and imaging makes histopathology interpretation a challenging task and is subjective relying on the expertise of pathologists. Additionally, the interpretation is laborious due to careful examination of multiples slides and multiple regions within the slide adding burden to pathologists. Computer aided diagnosis from histopathology slides will greatly assist the pathologists in their routine and will reduce the inter pathologist interpretation variability. An empirical study is proposed to investigate the proficiency of three prominent pre-trained networks -AlexNet, GoogLeNet and ResNet-50 in classifying the lung histopathology images belonging to three categories of lung pathology, lung adenocarcinoma, lung benign tissue and lung squamous cell carcinoma. The empirical study resulted in a phenomenal classification performance by all three classifiers with ResNet-50 model taking a marginal advantage over the other two classifiers with a classification accuracy of 99.26% and a Kappa coefficient of 0.989. AlexNet and GoogLeNet demonstrated a notable performance with 98.8% and 98.97% accuracy and a Kappa coefficient of 0.982 and 0.985.