Dealing With Speckle Noise in Deep Neural Network Segmentation of Medical Ultrasound Images

Sammanfattning: Segmentation of ultrasonic images is a common task in healthcare that requires time and attention from healthcare professionals. Automation of medical image segmentation using deep learning solutions is fast growing field and has been shown to be capable of near human performance. Ultrasonic images suffer from low signal-to-noise ratio and speckle patterns, noise filtering is a common pre-processing step in non-deep learning image segmentation methods used to improve segmentation results. In this thesis the effect of speckle filtering of echocardiographic images in deep learning segmentation using U-Net is investigated. When trained with speckle reduced and despeckled datasets, a U-Net model with 0.5·106 trainable parameters saw an rage dice score improvement of +0.15 in the 17 out of 32 categories that were found to be statistically different compared to the same network trained with unfiltered images. The U-Net model with 1.9·106 trainable parameters saw a decrease in performance in only 5 out of 32 categories, and the U-Net model with 31·106 trainable parameters saw a decrease in performance in 10 out of 32 categories when trained with the speckle filtered datasets. No definite differences in performance between the use of speckle suppression and full speckle removal were observed. This result shows potential for speckle filtering to be used as a means to reduce the complexity required of deep learning models in ultrasound segmentation tasks. The use of the wavelet transform as a down- and up-sampling layer in U-Net was also investigated. The speckle patterns in ultrasonic images can contain information about the tissue. The wavelet transform is capable of lossless down- and up-sampling in contrast to the commonly used down-sampling methods, which could enable the network to make use textural information and improve segmentations. The U-Net modified with the wavelet transform shows slightly improved results when trained with despeckled datasets compared to the unfiltered dataset, suggesting that it was not capable of extracting any information from the speckle. The experiments with the wavelet transform were far from exhaustive and more research is needed for proper assessment.