Topological Data Analysis of Thoracic Radiographic Images shows Improved Radiomics-based Lung Tumor Histology Prediction

Abstract

Topological data analysis (TDA) provides unparalleled tools to capture local to global structural shape information in data. In particular, its main method under the name of persistent homology has found many recent successful applications to both supervised and unsupervised machine learning. Despite its recent gain in popularity, much of its potential for medical image analysis remains undiscovered. In this paper we explore the prominent learning problems on thoracic radiographic images of lung tumors to which persistent homology provides improvements over state-of-the-art radiomic-based learning. It turns out that the novel topological features well capture complementary information important for both 'benign vs. malignant' and 'adenocarcinoma vs. squamous cell carcinoma' tumor prediction, while contributing less consistently to 'small cell vs. non-small cell'---an interesting result in its own right. Furthermore, while radiomic features may be better at predicting malignancy scores assigned by expert radiologists based on visual inspection, it turns out that topological features may be better at predicting the more accurate tumor histology assessed through long-term radiology review, biopsy, surgical resection, progression or response.