The deep learning model combining CT image and clinicopathological information for predicting ALK fusion status and response to ALK-TKI therapy in non-small cell lung cancer patients.

Affiliation

Song Z(#)(1), Liu T(#)(2)(3), Shi L(#)(2)(3), Yu Z(4), Shen Q(2)(3), Xu M(2)(3), Huang Z(5), Cai Z(6), Wang W(1), Xu C(7), Sun J(8), Chen M(9).
Author information:
(1)Department of Clinical Trial, Cancer Hospital of the University of Chinese Academy of Sciences
(Zhejiang Cancer Hospital), Hangzhou, 310022, Zhejiang, China.
(2)Shanghai Key Laboratory of Artificial Intelligence for Medical Image and Knowledge Graph, Shanghai, 200336, China.
(3)YITU AI Research Institute for Healthcare, Hangzhou, 310000, Zhejiang, China.
(4)Department of Medical Oncology, 900th Hospital, Fuzhou, 350000, Fujian, China.
(5)Department of Medical Oncology, Fujian Cancer Hospital, Fuzhou, 350001, China.
(6)Institute of Immunology, Zhejiang University School of Medicine, Hangzhou, 310009, Zhejiang, China.
(7)Department of Pathology, Fujian Cancer Hospital, Fuzhou, 350001, China.
(8)Department of Radiology, Cancer Hospital of the University of Chinese Academy of Sciences
(Zhejiang Cancer Hospital), Hangzhou, 310022, Zhejiang, China.
(9)Department of Radiotherapy, Cancer Hospital of the University of Chinese Academy of Sciences
(Zhejiang Cancer Hospital), Hangzhou, 310022, Zhejiang, China. [Email]
(#)Contributed equally

Abstract

PURPOSE: This study aimed to investigate the deep learning model (DLM) combining computed tomography (CT) images and clinicopathological information for predicting anaplastic lymphoma kinase (ALK) fusion status in non-small cell lung cancer (NSCLC) patients. MATERIALS AND METHODS: Preoperative CT images, clinicopathological information as well as the ALK fusion status from 937 patients in three hospitals were retrospectively collected to train and validate the DLM for the prediction of ALK fusion status in tumors. Another cohort of patients (n = 91) received ALK tyrosine kinase inhibitor (TKI) treatment was also included to evaluate the value of the DLM in predicting the clinical outcomes of the patients. RESULTS: The performances of the DLM trained only by CT images in the primary and validation cohorts were AUC = 0.8046 (95% CI 0.7715-0.8378) and AUC = 0.7754 (95% CI 0.7199-0.8310), respectively, while the DLM trained by both CT images and clinicopathological information exhibited better performance for the prediction of ALK fusion status (AUC = 0.8540, 95% CI 0.8257-0.8823 in the primary cohort, p < 0.001; AUC = 0.8481, 95% CI 0.8036-0.8926 in the validation cohort, p < 0.001). In addition, the deep learning scores of the DLMs showed significant differences between the wild-type and ALK infusion tumors. In the ALK-target therapy cohort (n = 91), the patients predicted as ALK-positive by the DLM showed better performance of progression-free survival than the patients predicted as ALK-negative (16.8 vs. 7.5 months, p = 0.010). CONCLUSION: Our findings showed that the DLM trained by both CT images and clinicopathological information could effectively predict the ALK fusion status and treatment responses of patients. For the small size of the ALK-target therapy cohort, larger data sets would be collected to further validate the performance of the model for predicting the response to ALK-TKI treatment.