Session Item

To exhaustively test and compare the performance of several models based either on radiomic or on deep learning approaches for the prediction of the 2-year overall survival (OS) in patients with non-small cell lung cancer (NSCLC).

The chest CT examinations of the 417 NSCLC patients included in the public LUNG1 dataset were used in the study.

For the radiomic approach, handcrafted features, extracted from the whole 3D tumour volume, were fed to 24 different pipelines formed by the combination of 4 feature selectors/reducers (i.e., ANOVA f-value, random forest, principal component analysis and feature agglomeration) and 6 classifiers (i.e., support vector machines, feed forward neural networks, nearest neighbours, bagging, random forest and extreme gradient boosting) to predict the 2-year OS. 

For deep learning, both the deep features extracted from the 2D tumour slices by a convolutional auto-encoder and the inner features learnt by a pre-trained convolutional neural network (CNN) were fed to the same 24 pipelines. In addition, the direct classification of the images with 3 different CNN architectures was tested, by considering both the original CT images and the synthetic images generated with a common data augmentation technique. The classification workflow and the total number of pipelines implemented for the three approaches are depicted in Figure 1. Finally, for each pipeline and approach, the performance with and without the inclusion of clinical parameters within the feature set was also evaluated in a cross-validation scheme.

For radiomics, the best pipeline was formed by the combination of feature agglomeration method and bagging classifier, which achieved an area under the receiver operating characteristic curve (AUC) of 0.683 and 0.652 on the training and test set, respectively. In general, deep learning approaches performed slightly worse than radiomics, both in terms of AUC and of difference between training and test performance. For the deep feature approach, the AUC of the best model was 0.692 and 0.631 on the training and test set, respectively, while for the CNN approach the greatest AUC was 0.692 and 0.644, respectively. The reduced predictive capabilities of deep learning methods could be ascribed to the bidimensional restriction of the analysis - necessary to reduce model complexity and increase the dataset size – which entails a poorer description of the tumour masses.

Using the public LUNG1 dataset, we extensively investigated and compared the performance of several classification pipelines based either on radiomic or on deep learning approaches for the prediction of the 2-year OS in NSCLC. The results of the best pipeline were comparable among the approaches and in line with previous works, demonstrating that these techniques are not able to extract additional information. However, we expect that the application of 3D deep learning models, in this work limited by the reduced size of the dataset, would increase the performance.