Neural network segmentation for deformation-induced martensitic transformation in X-ray computed tomograms

A U-Net-based neural network was employed to perform segmentation and binarization of X-ray computed tomography (CT) images containing austenite phase, martensite phase, and microvoids, which were degraded by background noise and imaging artifacts. We also identified the amount of training data required for precise segmentation and binarization of the martensite phase. The neural network achieved accurate segmentation for a base dataset comprising 100 training images, enabling reliable identification of each phase and microvoid. For images with higher contrast, accurate segmentation was achieved by incorporating a limited number of additional training images into the base dataset. For images containing a significantly larger fraction of the martensite phase, the trained network also yielded an appropriate segmentation result, although with reduced performance compared with the base training condition. Compared with conventional binarization methods, including Otsu’s global thresholding and adaptive thresholding, the neural network more accurately identified the martensite phase while effectively suppressing background noise and artifacts. Using a neural network trained with a limited amount of training data and training epochs, the three-dimensional (3D) deformation-induced martensitic transformation process was successfully visualized and tracked by extracting the martensite phase from noisy CT images.