Posted: Friday, February 4, 2022
A study conducted by Guotian Pei, MD, of Beijing Haidian Hospital, Beijing, China, and colleagues attempted to develop radiomic models for histology evaluation using CT images from patients with stage IA lung adenocarcinoma. These results demonstrated that radiomic models have the potential to predict histologic subtypes, suggesting it may serve as a noninvasive histology evaluation approach to the diagnosis and management of early lung adenocarcinoma. These findings were presented during the 2022 Annual Meeting of The Society of Thoracic Surgeons (STS).
CT images were collected from 290 patients with histologically diagnosed, stage IA lung adenocarcinoma who underwent complete resection. Tumors were defined on the InferScholar platform and labeled as minimally invasive adenocarcinoma (n = 91), invasive mucinous (n = 13) or nonmucinous (n = 163) adenocarcinoma, or precursor glandular lesions (n = 23).
A total of 1,454 features were extracted from these tumors, with 319 features retained after examining feature redundancy by Pearson correlation coefficient (< 0.8). Via principal component analysis, 40 features of varying categories were selected to create radiomic models for classification. Overall, the classification accuracy on the test set was 0.817.
The areas under the curve for categorizing precursor glandular lesions, minimally invasive adenocarcinoma, invasive nonmucinous adenocarcinoma, and invasive mucinous adenocarcinoma were 0.927, 0.882, 0.874, and 0.480, respectively; sensitivity and precision for differentiating precursor glandular lesions, minimally invasive adenocarcinoma, and invasive nonmucinous adenocarcinoma were 0.600 and 0.750, 0.842 and 0.800, and 0.909 and 0.857, respectively. Although these results indicate the decent capacity of the radiomics model to evaluate histology, the assessment of invasive mucinous adenocarcinoma should still be performed to improve the accuracy of this model, according to the investigators.
Disclosure: For disclosures of the study authors, visit eventscribe.net.