Growth Pattern Fingerprinting for Automatic Analysis of Lung Adenocarcinoma Overall Survival

Abstract

Lung adenocarcinoma (LUAD) tumour tissue grows into variable morphological architecture called growth patterns (GPs). The GPs are clinically linked to the biological behaviour of the tumour. However, due to the complex heterogeneity of the tumours, there is high inter-and intra-observer variability in the pathologist reporting of GPs. This paper proposes a deep-learning model for automatically classifying the LUAD growth patterns in whole slide images (WSIs). The model is trained and tested on 78 cases of LUAD in the digitised WSI of the sample. For each case, all the growth patterns were automatically classified and quantified. Our multivariate analysis shows that lepidic and micropapillary patterns are independent predictors for five-year survival (p < 0.05). The proposed model splits our study cohort into short- and long-term survival with p=0.009.

Keywords

• Digital pathology
• machine learning
• deep learning
• whole-slide image
• growth pattern
• survival analysis