Objective:

To investigate the method of malignancy prediction of pulmonary nodules based on radiomics.

Methods:

A total of 2 803 computed tomography (CT) images containing pulmonary nodules were extracted from 604 scans in the publicly available dataset of The Lung Image Database Consortium (LIDC)-Image Database Resource Initiative (IDRI). Each contour of nodules was labelled by the clinical doctor. Totally 96 high throughput features including gray level features

shape features and texture features were extracted according to the pulmonary nodule diagnosis criteria and put into the multi-class classifier based on the random forest to predict the malign

ancy. The degree of malignancy was classified into 1 to 5 levels. Among all images

1 000 of them were randomly chosen as the training set and the rest were used as the testing set. The experiment was repeated 10 times.

Results:

For a single nodule

the average prediction accuracy of five levels was 77.85%. The area under curve (AUC) of each category reached over 0.94. For each patient

the malignancy prediction accuracy of pulmonary nodules was 75.16%.

Conclusion:

The method of malignancy prediction of pulmonary nodules based on radiomics has a good performance. The results can provide a reliable basis for clinical diagnosis and help to detect the disease in the early stage.