Tissue-border Detection in Volumetric Laser Endomicroscopy using Bi-directional Gated Recurrent Neural Networks

Volumetric Laser Endomicroscopy (VLE) is a novel technique that can aid in the early detection of dysplasia in patients with Barrett's Esophagus (BE). Due to the relatively large size of VLE scans and the subtle appearance of this precursor of esophageal cancer, Computer Aided Detection (CAD) has proven to significantly contribute to VLE analysis. However, the tissue segmentation stage that is required for further processing by such a CAD system is currently limiting the execution speed, thereby hampering its clinical application. To segment the tissue of interest, current state-of-the-art solutions typically apply a Convolutional Neural Network (CNN) for image segmentation. In this paper, alternatively, a Recurrent Neural Network (RNN) is proposed to facilitate automatic tissue segmentation. In contrast to CNNs for image segmentation, the RNN focuses exclusively on the boundary between tissue and non-tissue rather than generating a full 2D segmentation, where we exploit the recent observation that only the upper tissue-boundary is relevant. This effectively poses the 2D segmentation problem as a 1D curve fitting problem, which can be computed much more efficiently. To train and evaluate our approach, four assessors manually annotated the tissue boundary in VLE images. The results of the proposed approach demonstrate a promising Mean Absolute Error (MAE) of 11 pixels or lower compared to human annotations, for which the inter-observer variability is around 5 pixels. Furthermore, the proposed RNN-based model is less computationally expensive compared to the current state-of-the-art model, since it requires approximately 18 times less floating-point operations. The resulting speed-up facilitates much faster pre-processing, enabling real-time CAD applications for VLE.