Quantifying Deep Learning Algorithm Performance for CT Image Segmentation of Composite Material Damage and Microstructures

Abstract

Deep Learning Algorithms (DLAs), particularly Convolutional Neural Networks (CNNs), have gained popularity for image segmentation of computed tomography (CT) scans because modern CNN architectures like U-Net remain effective even when trained on heavily augmented data. This significantly reduces human workload, requiring only a small fraction of the original dataset to be segmented manually. However, adoption for composite damage and microstructure segmentation tasks remains limited due to a lack of standardization in training methodology and accuracy reporting. In this work, statistical performance metrics from binary classification theory are used alongside application-specific metrics to quantify the accuracy of DLA image segmentations and illustrate the effects of image segmentation errors on the analysis of composite materials. These metrics revealed opportunities for improving segmentation performance through optimized training data selection and CNN architecture choice. Since small training datasets often poorly represent their parent datasets, a method for selecting representative training data using an image similarity algorithm is proposed. Each image is scored on its similarity to the overall dataset, and the local extrema are used to select images with common and unique features. This method significantly improved 2D U-Net’s segmentation accuracy of impact damage in carbon fiber reinforced polymer (CFRP) composites when given minimal training data. Architectural comparisons of CNNs revealed that 2.5D FC-DenseNet statistically significantly outperformed 2.5D U-Net for generalized CFRP impact damage segmentation tasks but took 52% longer to train. For ceramic matrix composite (CMC) microstructure image segmentation tasks, FC-DenseNet statistically matched U-Net’s accuracy while training 48% faster but was more susceptible to noise and artifacts based on qualitative analysis. These proposed methods and insights support ongoing efforts to optimize image segmentation workflows and automate future composite material characterizations.