Jionghui Gu1,2, Tong Tong2,3, Chang He1, Min Xu1, Xin Yang2,3, Jie Tian2,3,4, Tianan Jiang5,6, Kun Wang7,8. 1. Department of Ultrasound, The First Affiliated Hospital, College of Medicine, Zhejiang University, Hangzhou, 310003, China. 2. CAS Key Laboratory of Molecular Imaging, The State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China. 3. School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 100049, China. 4. Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, School of Medicine and Engineering, Beihang University, Beijing, 100191, China. 5. Department of Ultrasound, The First Affiliated Hospital, College of Medicine, Zhejiang University, Hangzhou, 310003, China. tiananjiang@zju.edu.cn. 6. Zhejiang Provincial Key Laboratory of Pulsed Electric Field Technology for Medical Transformation, Hangzhou, Zhejiang, China. tiananjiang@zju.edu.cn. 7. CAS Key Laboratory of Molecular Imaging, The State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China. kun.wang@ia.ac.cn. 8. School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 100049, China. kun.wang@ia.ac.cn.
Abstract
OBJECTIVES: Breast cancer (BC) is the most common cancer in women worldwide, and neoadjuvant chemotherapy (NAC) is considered the standard of treatment for most patients with BC. However, response rates to NAC vary among patients, which leads to delays in appropriate treatment and affects the prognosis for patients who ineffectively respond to NAC. This study aimed to investigate the feasibility of deep learning radiomics (DLR) in the prediction of NAC response at an early stage. METHODS: In total, 168 patients with clinicopathologically confirmed BC were enrolled in this prospective study, from March 2016 to December 2020. All patients completed NAC treatment and underwent ultrasonography (US) at three time points (before NAC, after the second course, and after the fourth course). We developed two DLR models, DLR-2 and DLR-4, for predicting responses after the second and fourth courses of NAC. Furthermore, a novel deep learning radiomics pipeline (DLRP) was proposed for stepwise prediction of response at different time points of NAC administration. RESULTS: In the validation cohort, DLR-2 achieved an AUC of 0.812 (95% CI: 0.770-0.851) with an NPV of 83.3% (95% CI: 76.5-89.6). DLR-4 achieved an AUC of 0.937 (95% CI: 0.913-0.955) with a specificity of 90.5% (95% CI: 86.3-94.2). Moreover, 19 of 21 non-response patients were successfully identified by DLRP, suggesting that they could benefit from treatment strategy adjustment at an early stage of NAC. CONCLUSIONS: The proposed DLRP strategy holds promise for effectively predicting NAC response at its early stage for BC patients. KEY POINTS: • We proposed two novel deep learning radiomics (DLR) models to predict response to neoadjuvant chemotherapy (NAC) in breast cancer (BC) patients based on US images at different NAC time points. • Combining two DLR models, a deep learning radiomics pipeline (DLRP) was proposed for stepwise prediction of response to NAC. • The DLRP may provide BC patients and physicians with an effective and feasible tool to predict response to NAC at an early stage and to determine further personalized treatment options.
OBJECTIVES: Breast cancer (BC) is the most common cancer in women worldwide, and neoadjuvant chemotherapy (NAC) is considered the standard of treatment for most patients with BC. However, response rates to NAC vary among patients, which leads to delays in appropriate treatment and affects the prognosis for patients who ineffectively respond to NAC. This study aimed to investigate the feasibility of deep learning radiomics (DLR) in the prediction of NAC response at an early stage. METHODS: In total, 168 patients with clinicopathologically confirmed BC were enrolled in this prospective study, from March 2016 to December 2020. All patients completed NAC treatment and underwent ultrasonography (US) at three time points (before NAC, after the second course, and after the fourth course). We developed two DLR models, DLR-2 and DLR-4, for predicting responses after the second and fourth courses of NAC. Furthermore, a novel deep learning radiomics pipeline (DLRP) was proposed for stepwise prediction of response at different time points of NAC administration. RESULTS: In the validation cohort, DLR-2 achieved an AUC of 0.812 (95% CI: 0.770-0.851) with an NPV of 83.3% (95% CI: 76.5-89.6). DLR-4 achieved an AUC of 0.937 (95% CI: 0.913-0.955) with a specificity of 90.5% (95% CI: 86.3-94.2). Moreover, 19 of 21 non-response patients were successfully identified by DLRP, suggesting that they could benefit from treatment strategy adjustment at an early stage of NAC. CONCLUSIONS: The proposed DLRP strategy holds promise for effectively predicting NAC response at its early stage for BC patients. KEY POINTS: • We proposed two novel deep learning radiomics (DLR) models to predict response to neoadjuvant chemotherapy (NAC) in breast cancer (BC) patients based on US images at different NAC time points. • Combining two DLR models, a deep learning radiomics pipeline (DLRP) was proposed for stepwise prediction of response to NAC. • The DLRP may provide BC patients and physicians with an effective and feasible tool to predict response to NAC at an early stage and to determine further personalized treatment options.