Hand-crafted and deep learning-based radiomics models for recurrence prediction of non-small cells lung cancers

Abstract

This research was created to examine the recurrence of non-small lung cancer (NSCLC) using computed-tomography images (CT-images) to avoid biopsy from patients because the cancer cells may have an uneven distribution which can lead to the investigation mistake. This work presents a comparison of the operations of two different methods: Hand-Crafted Radiomics model and deep learning-based radiomics model using 88 patient samples from open-access dataset of non-small cell lung cancer in The Cancer Imaging Archive (TCIA) Public Access. In Hand-Crafted Radiomics Models, the pattern of NSCLC CT-images was analyzed in various statistics as radiomics features. The radiomics features associated with recurrence are selected through three statistical calculations: LASSO, Chi-2, and ANOVA. Then, those selected radiomics features were processed using different models. In the Deep Learning-based Radiomics Model, the proposed artificial neural network has been used to enhance the recurrence prediction. The Hand-Crafted Radiomics Model with non-selected, Lasso, Chi-2, and ANOVA, give the following results: 76.56% (AUC 0.6361), 76.83% (AUC 0.6375), 78.64% (AUC 0.6778), and 78.17% (AUC 0.6556), respectively, and the Deep Learning-based Radiomic Models, including ResNet50 and DenseNet121 give the following results: 79.00% (AUC 0.6714), and 79.31% (AUC 0.6712), respectively.