Ultrasound Image Segmentation

My experiment extends the research in two directions. First, I propose a classical machine learning baseline: instead of directly predicting a segmentation mask, I perform feature extraction on the ultrasound input images and use classical regression models to predict the five ellipse parameters needed to represent the target anatomy: center x, center y , major axis, minor axis, and rotation angle. This approach tests whether the problem can be approximated as a structured geometric regression task, rather than full pixel-wise segmentation. The classical regressors serve as a baseline against which neural segmentation models can be compared in terms of both accuracy and interpretability.

Second, I replicate and extend the article’s neural network idea by comparing four related encoder–decoder architectures of increasing complexity: