Application of Machine Learning based on Habitat Imaging and Vision Transformer to Predict Treatment Response of Locally Advanced Esophageal Squamous Cell Carcinoma Following Neoadjuvant Chemoimmunotherapy : A Multi-center Study

Shuhan Xie¹Hai Zhang²Jinxin Xu³Shi-Jie Huang⁴Wen-Yi Liu⁵Zi-Lu Tang⁶Rong-Yu Xu⁶Sun-Kui Ke^3*Jin-Biao Xie^4*Qing-Yi Feng^2*Kang Mingqiang^1*

• ¹Fujian Medical University Union Hospital, Fuzhou, China
• ²The People’s Hospital of Gaozhou, Gaozhou, China
• ³Zhongshan Hospital, Xiamen University, Xiamen, Fujian Province, China
• ⁴Affiliated Hospital of Putian University, Putian, Fujian Province, China
• ⁵Cancer Hospital Shenzhen Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Shenzhen, China
• ⁶Quanzhou First Hospital, Fujian Medical University, Quanzhou, Fujian Province, China

Objective: Current medical examinations and biomarkers struggle to assess the efficacy of chemoimmunotherapy (nICT) for locally advanced esophageal squamous cell carcinoma (ESCC). This study aimed to develop a machine learning model integrating habitat imaging and deep learning (DL) to predict the treatment response of ESCC patients to nICT.The study retrospectively collected 309 ESCC patients from 6 medical centers, divided into training and external validation cohorts. For habitat imaging analysis, intratumoral subregions were clustered using the K-means clustering method.DL features from intratumoral and peritumoral subregions were extracted by Vision Transformer (ViT) respectively and then subjected to feature selection. Subsequently, 11 machine learning models were constructed for predictive model. The model's performance was evaluated using the area under the curve (AUC), decision curve analysis (DCA), calibration curve, and accuracy.Results: A total of 18 DL features were selected. The model of ExtraTrees, which was optimal, demonstrated superior performance with AUCs of 0.917 in training cohort and 0.831 in external validation cohort. Similarly, ExtraTrees showed good predictive capabilities in patients undergoing 2 cycles of nICT with AUC of 0.862 in validation cohort. This model also showed good calibration for prediction probability and satisfied clinical value on DCAs. Finally, the Shapley additive explanations method elucidated the model's precise predictions.The ExtraTrees model leveraging habitat imaging and ViT offered a non-invasive and accurate method to predict pathological response to nICT, guiding personalized treatment strategies, and decreasing the risk of immune-related adverse effects.