Radiographical assessment of tumour stroma and treatment outcomes using deep learning: a retrospective, multicohort study

Abstract

The tumour stroma microenvironment plays an important part in disease progression and its composition can influence treatment response and outcomes. Histological evaluation of tumour stroma is limited by access to tissue, spatial heterogeneity, and temporal evolution. We aimed to develop a radiological signature for non-invasive assessment of tumour stroma and treatment outcomes. In this multicentre, retrospective study, we analysed CT images and outcome data of 2209 patients with resected gastric cancer from five independent cohorts recruited from two centres (Nanfang Hospital of Southern Medical University [Guangzhou, China] and Sun Yat-sen University Cancer Center [Guangzhou, China]). Patients with histologically confirmed gastric cancer, at least 15 lymph nodes harvested, preoperative abdominal CT available, and complete clinicopathological and follow-up data were eligible for inclusion. Tumour tissue was collected for patients in the training cohort (321 patients), internal validation cohort one (246 patients), and external validation cohort one (128 patients). Four stroma classes were defined according to the protein expression of α-smooth muscle actin and periostin assessed by immunohistochemistry. The primary objective was to predict the histologically based stroma classes by using preoperative CT images. We trained a deep convolutional neural network model using the training cohort and tested the model in the internal and external validation cohort one. We evaluated the model's association with prognosis in the training cohort, two internal, and two external validation cohorts and compared outcomes of patients who received or did not receive adjuvant chemotherapy. The deep-learning model achieved a high diagnostic accuracy for assessing tumour stroma in both internal validation cohort one (area under the receiver operating characteristic curve [AUC] 0·96-0·98]) and external validation cohort one (AUC 0·89-0·94). The stromal imaging signature was significantly associated with disease-free survival and overall survival in all cohorts (p<0·0001). The predicted stroma classes remained an independent prognostic factor adjusting for clinicopathological variables including tumour size, stage, differentiation, and Lauren histology. In patients with stage II or III disease in predicted stroma classes one and two subgroups, patients who received adjuvant chemotherapy had improved survival compared with those who did not (in those with stage II disease hazard ratio [HR] 0·48 [95% CI 0·29-0·77], p=0·0021; and in those with stage III disease HR 0·70 [0·57-0·85], p=0·00042). However, in the other two subgroups adjuvant chemotherapy was not associated with survival and might even be detrimental in the predicted stroma class 4 subgroup (HR 1·48 [1·08-2·03], p=0·013). The deep-learning model could allow for accurate and non-invasive evaluation of tumour stroma from CT images in gastric cancer. The radiographical model predicted chemotherapy outcomes and could be used in combination with clinicopathological criteria to refine prognosis and inform treatment decisions of patients with gastric cancer. None.