Efficient Upper Limb Bone Region Extraction using Bi-directional Convolution

Abstract

The semantic segmentation of bone structures requires pixel-by-pixel classification and high extraction accuracy for objects in the image to build accurate bone models that can be used in diagnoses. Many segmentation methods have been developed, but the most common ones are based on Convolutional Neuronal Networks (CNNs). However, it has been reported that CNN-based segmentation methods cannot extract objects with complex shapes, such as a wrist, with high accuracy. One reason for this is the failure to consider the three-dimensional structure of medical images. Moreover, 3D-CNN methods have been proposed to tackle this problem, but 3D-CNNs require a huge amount of learning data. Considering bone segmentation, we should improve 2D-CNN models to apply to practical uses easily. Therefore, we propose a 2D-CNN-based segmentation method that uses bidirectional convolution processing and reconstructed images to take into account the three-dimensional structure of the bones in the upper limb region. Specifically, our method analyses images from two directions, axial and sagittal, with two different models combining BiConvLSTM and Attention U-Net. The images reconstructed from the sagittal plane to the axial plane are then used to obtain the attention map of the segmentation target. Performance experiments show that the proposed method exhibits an IoU of 0.9355, which is higher than U-Net and the state-of-the-method.