Lung cancer histology classification from CT images based on radiomics and deep learning models

• Published in: Medical & biological engineering & computing 2021, Vol.59 (1), p.215-226
• Main Authors: Marentakis, Panagiotis, Karaiskos, Pantelis, Kouloulias, Vassilis, Kelekis, Nikolaos, Argentos, Stylianos, Oikonomopoulos, Nikolaos, Loukas, Constantinos
• Format: Article
• Language: English
• Subjects: Adenocarcinoma; Artificial neural networks; Biomedical and Life Sciences; Biomedical Engineering and Bioengineering; Biomedicine; Classification; Clinical decision making; Combinatorial analysis; Computed tomography; Computer Applications; Decision making; Deep learning; Feature extraction; Histology; Human Physiology; Image classification; Image segmentation; Imaging; Long short-term memory; Lung cancer; Medical imaging; Neural networks; Non-small cell lung carcinoma; Original Article; Radiology; Radiomics; Small cell lung carcinoma; Squamous cell carcinoma; Transfer learning; Tumors
• ISSN: 0140-0118; 1741-0444
• DOI: 10.1007/s11517-020-02302-w

Adenocarcinoma (AC) and squamous cell carcinoma (SCC) are frequent reported cases of non-small cell lung cancer (NSCLC), responsible for a large fraction of cancer deaths worldwide. In this study, we aim to investigate the potential of NSCLC histology classification into AC and SCC by applying different feature extraction and classification techniques on pre-treatment CT images. The employed image dataset (102 patients) was taken from the publicly available cancer imaging archive collection (TCIA). We investigated four different families of techniques: (a) radiomics with two classifiers (kNN and SVM), (b) four state-of-the-art convolutional neural networks (CNNs) with transfer learning and fine tuning (Alexnet, ResNet101, Inceptionv3 and InceptionResnetv2), (c) a CNN combined with a long short-term memory (LSTM) network to fuse information about the spatial coherency of tumor’s CT slices, and (d) combinatorial models (LSTM + CNN + radiomics). In addition, the CT images were independently evaluated by two expert radiologists. Our results showed that the best CNN was Inception (accuracy = 0.67, auc = 0.74). LSTM + Inception yielded superior performance than all other methods (accuracy = 0.74, auc = 0.78). Moreover, LSTM + Inception outperformed experts by 7–25% ( p < 0.05). The proposed methodology does not require detailed segmentation of the tumor region and it may be used in conjunction with radiological findings to improve clinical decision-making. Graphical abstract Lung cancer histology classification from CT images based on CNN + LSTM.