Baseline whole-lung CT features deriving from deep learning and radiomics: prediction of benign and malignant pulmonary ground-glass nodules

Abstract

Objective: To develop and validate the model for predicting benign and malignant ground-glass nodules (GGNs) based on the whole-lung baseline CT features deriving from deep learning and radiomics.

Methods: This retrospective study included 385 GGNs from 3 hospitals, confirmed by pathology. We used 239 GGNs from Hospital 1 as the training and internal validation set; 115 and 31 GGNs from Hospital 2 and Hospital 3 as the external test sets 1 and 2, respectively. An additional 32 stable GGNs from Hospital 3 with more than five years of follow-up were used as the external test set 3. We evaluated clinical and morphological features of GGNs at baseline chest CT and extracted the whole-lung radiomics features simultaneously. Besides, baseline whole-lung CT image features are further assisted and extracted using the convolutional neural network. We used the back-propagation neural network to construct five prediction models based on different collocations of the features used for training. The area under the receiver operator characteristic curve (AUC) was used to compare the prediction performance among the five models. The Delong test was used to compare the differences in AUC between models pairwise.

Results: The model integrated clinical-morphological features, whole-lung radiomic features, and whole-lung image features (CMRI) performed best among the five models, and achieved the highest AUC in the internal validation set, external test set 1, and external test set 2, which were 0.886 (95% CI: 0.841-0.921), 0.830 (95%CI: 0.749-0.893) and 0.879 (95%CI: 0.712-0.968), respectively. In the above three sets, the differences in AUC between the CMRI model and other models were significant (all P < 0.05). Moreover, the accuracy of the CMRI model in the external test set 3 was 96.88%.

Conclusion: The baseline whole-lung CT features were feasible to predict the benign and malignant of GGNs, which is helpful for more refined management of GGNs.

Keywords: X-ray computed; deep learning; ground-glass nodules; lung cancer; radiomics; tomography.

Figure 1

The inclusion and allocation of patients with pathological results.

Figure 2

The inclusion and allocation of patients without pathological results.

Figure 3

The overall workflow of this study The white square of CNN is the first-order neural network based on the whole-lung image features; the white block of BPNN is the first-order neural network based on clinical-morphological features and whole-lung radiomics features. CNN: convolutional neural network, BPNN: back-propagation neural network, CM: the model based on clinical-morphological features, CMR: the model combined clinical-morphological features and whole-lung radiomics features, CMI: the model combined clinical-morphological features and whole-lung image features, CMRI: the model integrated clinical-morphological features, whole-lung radiomics features, and whole-lung image features, WR: the model based on whole-lung radiomics features.

Figure 4

Performance of different models in the prediction of benign and malignant GGN in sets with pathologically confirmed GGNs The ROC curves of five different models in each set are shown in the figure: (A) internal validation set, (B) external test set 1, and (C) external test set 2.

Figure 5

A case of malignant GGN was predicted successfully by the CMRI model The nodule was from the external test set 2. (A) A 69-year-old male presented with a small pGGN in the right upper lobe on baseline CT scan(white arrow). (B) The first review was performed after 293 days of follow-up and the lesion was slightly enlarged(white arrow). (C) A second examination was performed 691 days after follow-up, and the lesion was significantly enlarged and heterogeneous in density(white arrow). Sixteen days after the second review(for a total follow-up of 707 days), the nodule was surgically removed and pathologically confirmed the minimally invasive adenocarcinoma. (D–F) Heatmaps generated by GRAD-CAM for baseline, first review, and second review. Red or yellow areas represent high importance or strong activation, while blue or green areas indicate low importance or weak activation. The prediction scores of CM, WR, CMI, CMR and CMRI models were 0.667, 0.670, 0.718, 0.727 and 0.783, respectively. Compare these prediction scores with the threshold (0.764) calculated by the neural network: those with a value above the threshold were classified as malignant, and below were classified as benign. So, The CMRI model predicted this malignant nodule successfully based on the baseline CT features, whereas none of the CM, CMR, CMI, and WR models predicted correctly.

Figure 6

The long-term stable GGN that incorrectly predicted by the WR model The nodule was from the external test set 3 (without pathologically confirmed, all considered benign GGNs). (A, B) Are chest CT images of a 45-year-old female with a slice thickness of 1.5mm and 1.25mm, respectively. (A) Baseline CT showed a faint pGGN (white arrow) in the right upper lobe. (B) Follow-up CT of 2609 days (7.1 years) after baseline showed that the nodule was stable. This nodule was correctly predicted by four models other than the WR model. (C, D) Show the baseline and the follow-up heatmaps generated by GRAD-CAM, respectively. The prediction scores of CM, WR, CMI, CMR and CMRI models were 0.114, 0.799, 0.082, 0.103 and 0.094, respectively. Only the WR model had a prediction score above the threshold (0.764); therefore, This nodule was correctly predicted by four models other than the WR model. Although the nodule has not changed significantly after 7.1 years of follow-up, the heatmaps (D) activity is still increased compared with that of (C), which may indicate its slow progression.