Purpose
Borrmann classification in advanced gastric cancer (AGC) is necessarily associated with personalized surgical strategy and prognosis. But few radiomics research studies have focused on specific Borrmann classification, and there is yet no consensus regarding what machine learning methods should be the most effective.
Methods
A combined size of 889 AGC patients was retrospectively enrolled from two centers. Radiomic features were extracted from tumors manually delineated on preoperative computed tomography images. Two classification experiments (Borrmann I/II/III vs. IV and Borrmann II vs. III) were conducted. In each task, we combined three common feature selection methods and five typical machine learning classifiers to construct 15 basic classification models, and then fed the 15 predictions to a designed multilayer perceptron (MLP) network.
Results
In internal and external validation cohorts, the proposed ensemble MLP yielded good performance with area under curves of 0.767 and 0.702 for Borrmann I/II/III vs. IV, as well as 0.768 and 0.731 for Borrmann II vs. III. Considering the imbalanced distribution of four Borrmann types (I, 2.9%; II, 12.8%; III, 69.5%; IV, 14.7%), the ensemble MLP surpassed the overfitting barrier and attained fine specificity (0.667 and 0.750 for Borrmann I/II/III vs. IV; 0.714 and 0.620 for Borrmann II vs. III) and sensitivity (0.795 and 0.610 for Borrmann I/II/III vs. IV; 0.652 and 0.703 for Borrmann II vs. III). Also, survival analysis showed that patients could be significantly risk stratified by MLP predicted types in both experiments (p < 0.0001, log‐rank test).
Conclusions
This study proposed an MLP‐based ensemble learning architecture, which could identify Borrmann type IV automatically and improve the differentiation of Borrmann type II from III. The study provided a new view for specific Borrmann classification in clinical practice.