Radiomics and deep learning models to differentiate lung adenosquamous carcinoma: A multicenter trial
Xianjing Chu,
Lishui Niu,
Xianghui Yang,
Shiqi He,
Aixin Li,
Liu Chen,
Zhan Liang,
Di Jing,
Rongrong Zhou
Affiliations
Xianjing Chu
Department of Oncology, Xiangya Hospital, Central South University, Changsha 410008, China
Lishui Niu
Department of Oncology, Xiangya Hospital, Central South University, Changsha 410008, China
Xianghui Yang
Department of Oncology, Changsha Central Hospital, Changsha 410004, China
Shiqi He
Department of Computer Science, University of British Columbia, 2329 West Mall, Vancouver, British Columbia, Canada
Aixin Li
Department of Radiotherapy, The First Affiliated Hospital, Hengyang Medical School, University of South, Hengyang 421001, China
Liu Chen
Department of Oncology, Xiangya Hospital, Central South University, Changsha 410008, China
Zhan Liang
Department of Oncology, Xiangya Hospital, Central South University, Changsha 410008, China
Di Jing
Department of Oncology, Xiangya Hospital, Central South University, Changsha 410008, China; Corresponding author
Rongrong Zhou
Department of Oncology, Xiangya Hospital, Central South University, Changsha 410008, China; Xiangya Lung Cancer Center, Xiangya Hospital, Central South University, Changsha 410008, China; National Clinical Research Center for Geriatric Disorders, Xiangya Hospital, Central South University, Changsha 410008, China; Corresponding author
Summary: Adenosquamous carcinoma (ASC) is frequently misdiagnosed or overlooked in clinical practice due to its dual histological components and potential transformation from either adenocarcinoma (ADC) or squamous cell carcinoma (SCC). Our study aimed to differentiate ASC from ADC and SCC by incorporating features of enhanced CTs and clinical characteristics to build radiomics and deep learning models. The classification models were trained in Xiangya Hospital and validated in two other independent hospitals. The areas under the receiver operating characteristic curves (AUC), accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were used to estimate the performance. The optimal three-class classification model achieved a maximum AUC of 0.89 and accuracy of 0.81 in external validation sets, AUC of 0.99 and accuracy of 0.99 in the internal test set. These findings highlight the efficacy of our models in differentiating ASC, providing a non-invasive, timely, and accurate diagnostic approach before and during the treatment.
