A Comparative Study of Radiomics and Deep-Learning Based Methods for Pulmonary Nodule Malignancy Prediction in Low Dose CT Images

Abstract

ObjectivesBoth radiomics and deep learning methods have shown great promise in predicting lesion malignancy in various image-based oncology studies. However, it is still unclear which method to choose for a specific clinical problem given the access to the same amount of training data. In this study, we try to compare the performance of a series of carefully selected conventional radiomics methods, end-to-end deep learning models, and deep-feature based radiomics pipelines for pulmonary nodule malignancy prediction on an open database that consists of 1297 manually delineated lung nodules.MethodsConventional radiomics analysis was conducted by extracting standard handcrafted features from target nodule images. Several end-to-end deep classifier networks, including VGG, ResNet, DenseNet, and EfficientNet were employed to identify lung nodule malignancy as well. In addition to the baseline implementations, we also investigated the importance of feature selection and class balancing, as well as separating the features learned in the nodule target region and the background/context region. By pooling the radiomics and deep features together in a hybrid feature set, we investigated the compatibility of these two sets with respect to malignancy prediction.ResultsThe best baseline conventional radiomics model, deep learning model, and deep-feature based radiomics model achieved AUROC values (mean ± standard deviations) of 0.792 ± 0.025, 0.801 ± 0.018, and 0.817 ± 0.032, respectively through 5-fold cross-validation analyses. However, after trying out several optimization techniques, such as feature selection and data balancing, as well as adding context features, the corresponding best radiomics, end-to-end deep learning, and deep-feature based models achieved AUROC values of 0.921 ± 0.010, 0.824 ± 0.021, and 0.936 ± 0.011, respectively. We achieved the best prediction accuracy from the hybrid feature set (AUROC: 0.938 ± 0.010).ConclusionThe end-to-end deep-learning model outperforms conventional radiomics out of the box without much fine-tuning. On the other hand, fine-tuning the models lead to significant improvements in the prediction performance where the conventional and deep-feature based radiomics models achieved comparable results. The hybrid radiomics method seems to be the most promising model for lung nodule malignancy prediction in this comparative study.

Clinical radiomics-based machine learning versus three-dimension convolutional neural network analysis for differentiation of thymic epithelial tumors from other prevascular mediastinal tumors on chest computed tomography scan

“…Radiomics with ML and DL are active research in the field of oncology ( 17 ). Some studies showed that the DL model had better performance than the ML-based radiomics ( 18 – 20 ), some showed ML-based radiomics out-performed DL model ( 21 ), and some demonstrated DL-based radiomics model had the best performance ( 22 , 23 ). Prior studies had performed radiomics based ML or DL to classify thymoma form other PMT ( 24 – 26 ).…”

Application of PET/CT-based deep learning radiomics in head and neck cancer prognosis: a systematic review

“…Thus, the combination of hand-crafted features and deep features improves the classification or regression performance of the model [57]. Research on lung cancer by Wang [58], Afshar [59], Astaraki [60], and Liang [61]; on breast cancer by Jiang [62]; on Gastric cancer by Sun [63] and Dong [15]; and on glioma by Chen [64] has indicated that deep learning features are complementary to manual features, and that combining HCR and DLR provides more comprehensive features and enables the model to achieve better results. However, because deep learning features are automatically learned in a black box-like process, they are generally nameless [65].…”

A Review of Convolutional Neural Network Development in Computer Vision

“…The model first uses forward propagation in the convolutional layer to connect each layer with other layers in the network and then uses the feature maps of all previous layers as the input of each subsequent layer to construct DenseNet [79,80]. On popular image classification benchmarks, DenseNet can achieve comparable accuracy to ResNet in the ImageNet image classification competition, but it requires significantly fewer parameters [81,82]. In addition, it improves the gradient vanishing problem, and at the same time, it strengthens the feature propagation process and promotes feature reuse through the reorganization of feature maps, reducing the amount of irrelevant computation.…”