This study presents an enhanced methodology for bone image segmentation in Ultrasonic Computed Tomography (USCT) utilizing the Adamax optimizer. Our approach focuses on optimizing a deep-learning-based neural network architecture to achieve efficient and accurate automatic segmentation of bone images. Initially, we improve the Variable Structure Model of Neuron (VSMN) for effective USCT noise removal and data augmentation. Subsequently, we train and evaluate a VGG-SegNet neural network on previously unseen USCT images using the Adamax optimizer. This dual-phase process ensures robust noise reduction and precise segmentation. We provide an open-access USCT dataset to facilitate further research and validation. The model is implemented on both CPU and GPU, demonstrating significant performance improvements with training and validation accuracies of 97.38% and 96%, respectively, and a minimal segmentation error of 0.006. The Adamax optimizer enhances the network’s ability to handle the complexities of USCT data, leading to high segmentation accuracy and efficient processing times. Our method showcases superior performance compared to existing techniques, highlighting its potential for clinical applications in bone imaging. This work contributes to the advancement of medical imaging technologies by offering a reliable and effective solution for automatic bone segmentation in USCT.