Enhancing Classification of Indeterminate Pulmonary Nodules at Multiple Scans in NLST Dataset through Integrated Spectral CT Radiomics

Accurate classification of malignant pulmonary nodules in low-dose CT (LdCT) screening–detected nodules remains a clinical challenge, especially for nodules labeled by experts as indeterminate lesions (IDLs). Resolving IDL malignancy necessitates follow-up evaluations, which can be costly and carry variable risks. Existing machine learning (ML) algorithms have shown classification performance comparable to expert readers and face the same challenge of resolving IDL malignancy. In this study, we investigate a CT radiomics–machine learning (CT-omics-ML) pipeline to address this challenge using data from the National Lung Screening Trial (NLST) database. All IDLs included in this study have biopsy-confirmed malignancy and expert-validated lesion segmentation. The proposed CT-omics-ML pipeline extracts energy-sensitive tissue pathological characteristic (EsTPC) features from a series of virtual or effective monoenergetic images (EMIs) derived from conventional LdCT screening data. Multiple sets of EsTPC features are subsequently integrated via adaptive multi-energy redundancy reduction along the energy series to construct a compact and discriminative representation. The performance of CT-omics-ML pipeline was compared to two baselines of (1) conventional radiomic features, which were extracted from the conventional LdCT screening nodule images, called Radiomics hereafter, and (2) existing deep learning (DL) algorithms, which use convolutional neural network (CNN) to extract abstract features from the conventional LdCT screening nodule images and classify the abstract features for prediction of lesion malignancy, called DL-CNN hereafter. The baseline of Radiomics achieved AUC (area under the receiver operating characteristic curve) values in the range from 0.612 to 0.832, comparable to the expert performance – indeterminate. The baseline of DL-CNN achieved AUC values ranging from 0.600 to 0.784. The CT-omics-ML pipeline reached AUC values of 0.822 to 0.998. The gain is obvious. These results demonstrate that explicit modeling of energy-dependent tissue characteristics combined with adaptive multi-energy integration substantially improves malignancy discrimination over both baselines of Radiomics and DL-CNN.