Abstract:

Background and purpose: Tumor heterogeneity exists in gastric cancer. The expression of human epidermal growth factor receptor 2 (HER2) is significantly different in some gastric cancer patients before and after neoadjuvant chemotherapy (NAC). In order to further study the heterogeneity of HER2 expression and reduce the visual error of human interpretation, we used phase Ⅲ artificial intelligence (AI) to assess the HER2 status of gastric cancer patients pre-NAC and post-NAC to evaluate the practicability and feasibility of AI, and studied the effect of NAC on HER2 expression in gastric cancer patients, so as to provide reference for subsequent treatment. Methods: Clinical data of 397 gastric cancer patients receiving NAC were collected from a multicenter, prospective, randomized controlled phase Ⅲ clinical trial (NCT01516944) mainly based on on the Fourth Hospital of Hebei Medical University. HER2 expression in samples of pre-NAC endoscopic biopsies and post-NAC surgical specimens was first assessed visually by 2 pathologists using optical microscope followed by AI-assisted microscope respectively. The comprehensive results of three senior pathologists were set as the gold standard. The consistency between the two assessment results and the gold standard was analyzed, and the clinicopathological features affecting the expression changes of HER2 pre- and post-NAC and the effect of HER2 change on prognosis were also explored. Results: Compared with visual assessment, the consistency between AI assessment results and gold standard was better (0.766 vs 0.853, 0.773 vs 0.876). HER2 expression assessed by AI was down-regulated in 97 patients (24.43%) post-NAC and up-regulated in 27 patients (6.80%). Change of HER2 expression was significantly correlated with lymph node metastasis (ypN, P=0.019) and tumor regression grade (P=0.003). Poor tumor regression was an independent risk factor for the upregulation of HER2 (P=0.032). The 5-year overall survival (OS) rate and disease-free survival (DFS) rate of pathologic complete response (pCR) patients were significantly better compared with non-pCR patients (92.9% vs 42.5%, P=0.002; 92.9% vs 36.0%, P=0.001). In non-pCR patients, the 5-year OS and DFS rates of HER2 down-regulated patients post-NAC were better compared with HER2 up-regulated patients(56.7% vs 30.5%, P<0.001; 56.1% vs 23.0%, P<0.001). And for those HER2 2+/3+ non-pCR patients, the 5-year OS and DFS rates of HER2 down-regulated patients post-NAC were better compared with HER2 unchanged patients (56.7% vs 33.5%, P=0.003; 56.1% vs 32.0%, P=0.002). Conclusion: The application of AI technology in HER2 assessment of NAC patients with gastric cancer can reduce visual measurement error, and AI could be a powerful tool to assess HER2 expression efficiently and accurately. Gastric cancer patients with pCR and down-regulated HER2 post-NAC are more likely to achieve better long-term survival. If upregulation of HER2 status may suggest poor prognosis, attention should be paid closely to the recurrence and metastasis of such patients, and the treatment should be adjusted in time.

Key words: Gastric cancer, Neoadjuvant chemotherapy, Human epidermal growth factor receptor 2, Artificial intelligence