Prediction of Malignancy in Lung Cancer using several strategies for the fusion of Multi-Channel Pyradiomics Images

Radiomics has become a reference tool for the early diagnosis of cancer in screening programs. Radiomics involves the extraction of a large number of quantitative features from medical scans. Such features define a representation space of lesions that should capture the heterogeneity of tissue texture in order to discriminate malignacy.This work presents the process of generation and subsequent study of a representation space of pulmonary nodules given by the deep convolutional features of multiple texture images extracted from computed tomography (CT). The objective of the study is to assess whether this deep texture features have a positive impact on the diagnosis of lung cancer compared to deep intensity features. To do so, we have trained SVM models with different data splits, evaluating the diagnostic performance using metrics defined at, both, slice and nodule levels on an own data base for early diagnosis of small (<3cm) pulmonary nodule. Results show that deep texture features perform better than intensity ones with a raise in specificity from 0.49 to 1 at slice level and 0 to 0.67 at nodule level.