Incorporating spatial features into faster R-CNN for cancer detection in histopathological analysis

Abstract

Cancer is one of the most serious and common disease in the world, and stomach cancer is one of the most common cancer type. Early detection is important to alleviate risk of cancer. With the improvement of technology, digital pathology based on deep learning plays a role in early detection. There are two things to consider in histopathological analysis : one thing is cancer lesion has unclear boundary and various magnification levels should be reflected on spatial features. However, current studies have focused on channel based features with attention mechanism to treat unclear boundary or various magnification levels have not been applied into spatial features. In this study, we introduce an attention module composed of channel attention module and spatial attention module, and spatial attention module adopted local pooling for treating magnification level. Attention module is applied to our base model, Faster R-CNN. Experimental result shows our method shows better accuracy in quantitative and qualitative manner. We hope our method can contribute to make more reliable and accurate cancer detection model.