Abstract Background Radiomic features of perinodular regions make contribution to lung nodules classification. To date, researchers roughly dissected 5mm ring around the nodules, which may omit valuable information. Purpose To find most informative and valuable perinodular region and explore their potential in nodules classification. Materials and Methods We collected CT images of 6847 patients(malignant: 4750 patients; benign: 2097 patients;median: 54.4 years ) from three centers. We used two types of tumor amplification mask to obtain radiomic features from different region: fixed-distance mask and diameter-multiplication mask. To discriminate malignant and benign nodules, we input radiomic features from perinodular region into three types of models. Results A total of 256 features were learned by deep-learning models,256 features for radiomics model and 512 features for hybrid models. Then we used purely perinodular regions to discriminate benign and malignant nodules. The first method extracted the most information at 3 mm with an AUC of 0.88 while the second method extracted more stable information and the model performed best at 1.3 times with an AUC of 0.86 on test set (multiplicative: IQR:0.0141;fixed-distance:IQR:0.0213) . In addition, we used intra- and peri-tumor radiomic features to differentiate malignant nodules and subtypes of nodules. Radiomics model achieved an AUC of 0.915 on test set, Cam-ResNet achieved an AUC of 0.929 and hybrid models obtained highest AUC of 0.933. Conclusions The hybrid model could more accurately classify lung nodules, achieved highest AUC of 0.933 This abstract is funded by: This work was supported by Noncommunicable Chronic Diseases-National Science and Technology Major Project（2023ZD050610X / 2023ZD0506100 to W Li）
Bi et al. (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: