MILISS: Multi-instance Learning with Instance Selection Strategy for Gastric Cancer Classification

Accurate distinction of histopathological subtypes of gastric cancer is essential for individualized treatment. However, accurate gastric cancer subtypes classification remains challenging due to the chaotic and disordered distribution of gastric cancer-positive tissues, and the fact that mixed type organizations resemble all other subtypes, as well as the lack of instance-level labeled medical image data. To this end, a multi-instance learning gastric cancer subtypes method incorporating instance selection strategy is proposed. The method begins by constructing a public anchor set through the collection of key instances, next selects appropriate instances as inputs to the backbone network using the public anchor set as a benchmark, then aggregates the obtained instance-level features into slide-level features through one-dimensional maximum pooling, and finally sends them to the multi-instance classifier to obtain slide-level prediction results. Experimental results on the Lauren subtypes dataset of gastric cancer show that the method has higher classification accuracy than existing classification methods.