Segmentation of induced substantia nigra from transcranial ultrasound images using deep convolutional neural network

Abstract

Parkinson's disease (PD) is an age-related neurodegenerative disorder, whose early diagnosis is challenging. PD neuropathology is characterized by a selective loss of dopaminergic neurons in the substantia nigra (SN). The echogenicity of SN is considered as an important biomarker for diagnosing PD. Since Ultrasound is well suited for measuring echogenicity, transcranial ultrasound images (TCUI) are used to diagnose PD and have become an industrial standard. But, ultrasound images usually have low resolution and are noisy compared to other medical imaging modalities. Thus, this whole method relies on the experience of the clinician to identify SN from the TCUI. To automate the process, we propose a deep convolutional neural network based on the U-Net architecture with a weighted binary cross-entropy (WBCE) loss function to do semantic segmentation of SN from the TCUI obtained. To alleviate the negative effects caused by noisy images, third-order Volterra filter is used for pre-processing. Furthermore, we compare the convergence rate of WBCE with standard BCE and other standard loss metrics, where WBCE outperforms others. We then, compare the results of the proposed U-Net architecture with a DenseNet based architecture and a U-Net and DenseNet based hybrid architecture, where we obtain better accuracy with the proposed architecture.