Segmentation of Median Nerve by Signal Processing and Artificial Intelligence on Ultrasound Images

Abstract

In medical sciences, visualizing the internal dynamics and structures of the human body throughout bodily function is critical, and ultrasound imaging (US) is one of the most extensively utilized medical imaging technologies. Carpal tunnel syndrome (CTS) is a kind of peripheral neuropathy, a frequently occurring disease in the wrist that affects many people. When the median nerve is compressed within the carpal tunnel, it produces a variety of nerve function problems that manifest as CTS. In this study, automatic segmentation of median nerve using signal processing and convolutional neural network (CNN)-based methods were studied.

In signal processing, mathematical morphology, edge detection, and contouring are employed, while in convolutional neural network (CNN), U-Net is used. The performance of signal processing techniques was engineered by concentrating on structural or kernel alterations for the signal processing method. The base U-Net, U-Net with pre-processed data, and U-Net with augmented and pre-processed data with batch norm layer are the three architectures evaluated in deep learning. The dice score, accuracy, Jaccard Similarity coefficient, recall, precision, and F1 score are all used to compare the results.

The signal processing technique observed a significant correlation of cross-correlation coefficient (CSA) between the ground truth (GT) and the segmented image with a close resemblance of over 90% with a correlation coefficient of 0.962 when tested on the 35 images. However, the model has estimated the CSA of the median nerve as normal in several situations, even when the expert or sonographer evaluated it as abnormal. The process of feature extraction, however, is where this technique's shortcoming lies. It took more time and processing to manually modify the kernel's weight and iterate numerous times to segment the median nerve. Furthermore, the approach was not reliable and favored certain feature images over others.

The U-Net model trained with pre-processed data, and augmented data with batch norm layer surpassed the two other models and achieves amazing results in median nerve segmentation. When evaluated on test datasets, an accuracy of 99.8% was achieved, which is 14.1 % higher than approach one (base U-Net) and 4.4% higher than approach two (U-Net with pre-processed data). The method was also quite successful in finding the median nerve, with a dice similarity coefficient (DSC) of 0.899, which was much higher than the other two approaches. This shows that when deep learning is given additional training data and the input data is cleaned, the outcomes are more accurate. This implies that data pre-processing and data augmentation are important not just for cleaning data and expanding the number of datasets, but also for improving accuracy. This demonstrates that this model could be used as a screening tool in clinical practice to expedite the identification, diagnosis, and assessment of CTS.

-