Ultrasound Elastography using Machine Learning

This thesis aims at solving two main problems that we face in ultrasound elastography, namely fast strain estimation and radio frequency (RF) frame selection. We rely on machine learning concepts such as Principal Components Analysis (PCA), multi-layer perceptron (MLP) and convolutional neural networks (CNN) to build 3 models that are trained on both phantom and in vivo data. In our first work, we developed a method to estimate the initial displacement between two ultrasound RF frames using PCA. We first compute an initial displacement estimate of around 1% of the samples, and then decompose the displacement into a linear combination of principal components (obtained offline during the training step). Our method assumes that the initial displacement of the whole image could also be described by this linear combination of principal components. This yields the same result that we could have had if we run dynamic programming (DP). The advantage of using PCA is that we could compute the same initial displacement image more than 10 times faster than DP. We then pass the result to GLobal Ultrasound Elastography (GLUE) for fine-tuning it, so we call the method PCA-GLUE.
\n
\nIn our second work, we developed a novel method to address the problem of RF frame selection in ultrasound elastography. Intuitively, we would like to have a classifier that gives a binary 1 to
\nRF frame pairs that yield high-quality strain images. We make use of our previous work where we decompose the initial displacement between two RF frames into a weight vector multiplied by some principal components. We consider the weight vector as our input feature vector to an MLP model. Given two RF frames I1 and I2, the MLP model predicts the normalized cross correlation (NCC) between the two RF frames I1 and I2′ (I2′ is I2 after being displaced according to the displacement of GLUE/PCA-GLUE). Our final contribution in this line of research is the introduction of a CNN-based method for RF frame selection as follows. First, we changed the architecture from an MLP model to a CNN that takes the two RF frames on two channels. The CNN has better results compared to the MLP model due to having more features. Second, we improved the automatic labelling of the data by having physical conditions that must be satisfied together in order to consider the pair as a suitable pair of RF frames.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.