Real-time application based CNN architecture for automatic USCT bone image segmentation

Artificial Intelligence (AI) in medical image analysis has achieved excellent success in automatic diagnosis in the same way as clinician, especially in the ultrasound field. In this work, we develop a new segmentation application based on various Convolutional Neural Network (CNN) models for Ultrasonic Computed Tomographic (USCT) images. To evaluate the proposed segmentation system, we use different state-of-the-art models for better segmentation performances to train and test the suggested system. We ensure in this work a USCT data augmentation technique based on the Haar wavelet transform and the improved k-means algorithms. Thus, we offer a free dataset for USCT researchers. Moreover, the proposed CNN system is trained and tested using the networks of Adadelta and Adam optimizers. The whole system is implemented on a CPU and a GPU for complexity analysis. High segmentation accuracy has been achieved using the Adadelta optimizer, reaching 99.24%, 99.19%, 99.13% and 99.10% for VGG-Segnet, VGG-Unet, Fully CNN (FCN)-8 and FCN-32 models, respectively. To obtain better results, we use the Adam optimizer to train and test different architectures, and we obtain more competitive results attaining 99.55%, 99.31%, 99.35% and 99.45% for VGG-Segnet, VGG-Unet, FCN-8 and FCN-32, respectively. The achieved results outperform the state of the art in terms of accuracy and time speed up. Moreover, our proposed CNN segmentation confirms the low computational complexity of the system. In addition, our system proves to be a good candidate for medical real-time applications thanks to its implementation on the GPU.