DeepScribble: Interactive Pathology Image Segmentation Using Deep Neural Networks with Scribbles

Sungduk Cho, Hyungjoon Jang, Jing Wei Tan, Won-Ki Jeong

Abstract: Tumor segmentation is a challenging but crucial task in digital pathology for accurate diagnosis. Recent studies on deep neural networks have shown promising results in various image segmentation problems. However, unlike several medical imaging modalities (such as computerized tomography, CT and magnetic resonance imaging, MRI), the boundary between the normal and the tumor area in pathology images is usually fuzzy and ambiguous, making it difficult to adapt conventional image segmentation methods to these images. In this paper, we propose an interactive segmentation method that corrects the segmented boundaries from deep neural networks using user interaction. The proposed method functions in two stages; the first network initially generates the best prediction of the tumor boundary; the second network then refines the segmentation iteratively by exploiting the user scribble annotations on-the-fly. Our approach leverages the feature learning aspects of deep neural networks in the correction step to reduce user effort. We demonstrate the efficacy of the proposed method on real pathology images.