Exploring the predictive value of additional peritumoral regions based on deep learning and radiomics: A multicenter study

Purpose: The present study assessed the predictive value of peritumoral regions on three tumor tasks, and further explored the influence of peritumors with different sizes.

Methods: We retrospectively collected 333 samples of gastrointestinal stromal tumors from the Second Affiliated Hospital of Zhejiang University School of Medicine, and 183 samples of gastrointestinal stromal tumors from Tianjin Medical University Cancer Hospital. We also collected 211 samples of laryngeal carcinoma and 233 samples of nasopharyngeal carcinoma from the First Affiliated Hospital of Jinan University. The tasks of three tumor datasets were risk assessment (gastrointestinal stromal tumor), T3/T4 staging prediction (laryngeal carcinoma), and distant metastasis prediction (nasopharyngeal carcinoma), respectively. First, deep learning and radiomics were respectively used to construct peritumoral models, to study whether the peritumor had predictive value on three tumor datasets. Furthermore, we defined different sizes peritumors including fixed size (not considering tumor size) and adaptive size (according to average tumor radius) to explore the influence of peritumor of different sizes and types of tumors. Finally, we visualized the deep learning and radiomic models to observe the influence of the peritumor in three datasets.

Results: The performance of intra-peritumors are better than intratumors alone in three datasets. Specifically, the comparisons of area under receiver operating characteristic curve in the testing set between intra-peritumoral and intratumoral models are: 0.908 vs 0.873 (P value: 0.037) in gastrointestinal stromal tumor datasets, 0.796 vs 0.756 (P value: 0.188) in laryngeal carcinoma datasets and 0.660 vs 0.579 (P value: 0.431) in nasopharyngeal carcinoma datasets. Furthermore, for gastrointestinal stromal tumor datasets, deep learning is more stable to learn peritumors with both fixed and adaptive size than radiomics. For laryngeal carcinoma datasets, the intra-peritumoral radiomic model could make model performance more balanced. For nasopharyngeal carcinoma datasets, radiomics is also more suitable for modeling peritumors than deep learning. The size of the peritumor is critical in this task, and only the performance of 1.5 mm-4.5 mm peritumors is stable.

Conclusions: Our results indicate that peritumors have additional predictive value in three tumor datasets through deep learning or radiomics. The definitions of the peritumoral region and artificial intelligence method also have great influence on the performance of the peritumor.