An automatic B-snake model based on deep learning for medical image segmentation

Intensive research has been carried out on Active Contours (AC) models and many efforts have been made to increase their robustness. The success of an AC model relies heavily on the robustness of its energy. However, in the case of poor-quality images, the energy functional, which is based on contour information (pixel intensity changes), may be low, which can contribute to the divergence of the AC process. To tackle these issues, we introduce a new Automatic Deep learning-based Spline Active Contour (ADS-AC) approach. Our proposed AC method incorporates an energy term that distinguishes different textures within an image by combining two learning models (Auto-Encoder and U-Net). In addition, to ensure the obtained contour aligns precisely with the edges of the target object, we introduce an innovative adjustment step called Snake Adjustment Strategy (SAS), which operates in the balanced state of our snake. The robustness and efficiency of the proposed ADS-AC-SAS model is evaluated on three different datasets of brain MRI images, breast ultrasound images and Kvasir-SEG database. Results demonstrate that our model significantly enhances segmentation performance and exhibits robustness to noise and various challenges commonly encountered in medical images.