Integrating deep learning and radiomics in the differentiation of major histological subtypes of invasive non-mucinous lung adenocarcinoma using positron emission tomography and computed tomography.

Background: Accurate identification of the main subtypes of invasive non-mucinous adenocarcinoma (INMA) is essential for individualized treatment. Given the limitations of preoperative biopsy pathological diagnosis, this study aimed to evaluate the feasibility of using positron emission tomography and computed tomography (PET/CT) radiomics features, deep learning (DL) features, and their combined models for non-invasive identification of the four main INMA subtypes prior to surgery. Methods: A total of 386 patients from hospital one and 32 patients from hospital two with preoperative PET/CT images were enrolled as the training, internal validation cohorts and external validation cohorts, retrospectively. Radiomics features were extracted from CT, PET and PET/CT images to build radiomics models using various machine learning (ML) classifiers. DL features were extracted using a Resnet34 trained to extract DL features from the best performance model. Different fusion techniques were utilized to identify subtypes. Different fusion techniques were employed to integrate the radiomic and DL features to build a final fusion model for subtype differentiation. Results: The radiomics model using support vector machine (SVM) classifier achieved an area under the curve (AUC) of 0.79, accuracy of 0.70, and precision of 0.70 for differentiating the four INMA subtypes. The DL model demonstrated the best performance in the internal validation cohort with an AUC and accuracy of 0.89 and 0.74, respectively. The combined mode integrating radiomic and DL features from fused PET/CT images achieved the highest performance in the external validation cohort with an AUC, accuracy and precision of 0.85, 0.81, and 0.85, respectively. Conclusions: Integrating DL and radiomic features derived from PET/CT images is a feasible and accurate approach for differentiating the four main subtypes of INMA preoperatively.