Reproducing a segmentation baseline

Start with a reproducible baseline

In this student research project, Levi Walker and I studied a conventional 2D U-Net for brain-tumor segmentation. The purpose was to reproduce a baseline and examine targeted changes in a controlled setting.

From images to regions

Unlike classification, segmentation assigns a prediction to image locations. The work uses multiple MRI modalities and evaluates the overlap between predicted tumor regions and reference masks.

Evaluate the changes carefully

The poster compares the baseline with limited rotation augmentation and segmentation threshold selection on validation data. Repeated runs support the comparison of performance and training stability.

This is a research study on a specific dataset and setup. It does not establish clinical readiness or performance on other populations.