1210MO - Comparison of clinical and radiological features-based and radiomics features-based models for pulmonary nodule malignancy prediction on CT images

Background

Currently, early diagnosis of pulmonary nodules still relies on radiologist reading of CT images, which needs a heavy workload, extensive experiences, and is somewhat subjective. We propose an automatic radiomics-based approach to predict the risk of lung cancer.

Methods

A total of 1,895 patients with 1,909 surgical pathology confirmed 5-30 mm pulmonary nodules (1,181 malignant, 728 benign) were enrolled in 31 centers, with clinical information (gender, age) and chest CT scans. 25 radiological features were extracted by experienced radiologists. The regions of interest (ROIs) containing target nodules on the CT images were automatically segmented by a 3D U-net model, and 2,153 radiomics features were extracted using PyRadiomics. A clinical and radiological features-based model (CRFM) and a radiomics features-based model (RFM) for pulmonary nodule malignancy prediction were constructed in a randomly sampled training set (n=950), based on the average probabilities of Randomforest, LightGBM, and Lasso algorithms, and verified independently (n1=397, n2=562). Meanwhile, a combined model integrating the scores of the above two models was established (n=397) using logistic regression, and validated independently (n=562).

Results

In two validation sets (n1=397, n2=562), the CRFM model got AUCs of 0.912 (0.882-0.942) and 0.893 (0.864-0.921), and accuracies of 0.791 (0.748-0.828) and 0.794 (0.758-0.825), while RFM model obtained AUCs of 0.881 (0.847-0.915) and 0.863 (0.832-0.894), and accuracies of 0.781 (0.738-0.819) and 0.778 (0.741-0.810), respectively. The performance of RFM model was noninferior to CRFM model as evidenced by non-significant difference in AUCs (p>0.05) in both validation cohorts. However, the combined model improved diagnostic accuracy with elevated AUCs of 4.6% (vs. RFM, p＜0.001) and 1.6% (vs.CRFM, p=0.0196) in the independent validation set (n=562), respectively.

Conclusions

Radiomics features-based model for early diagnosis of pulmonary nodules is not inferior to clinical and radiological features-based model, with potential application when CT scan interpretation by experienced radiologists are unavailable.

Clinical trial identification

NCT03181490; NCT03651986.

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Guangdong Basic and Applied Basic Research Foundation (2020A1515011454), National Natural Science Foundation of China (No. 81902311), The Leading Projects of Guangzhou Municipal Health Sciences Foundation (No. 20211A01172).

Disclosure

B. Wang, M. Peng, J. Tao, X. Tu, H. Lu, X. Qiu, Y. Yang: Financial Interests, Personal, Full or part-time Employment: AnchorDX Medical Co., Ltd. All other authors have declared no conflicts of interests.