Overview
Developed with Monash Analysis of Images in Medicine (Monash AIM), this project segments stroke lesions in brain scans: for every pixel, the model decides whether it belongs to a lesion, producing a mask clinicians could use to see where the stroke is and how large it is.
Approach
The model is based on MSEG-HarDNet, a segmentation network built on HarDNet (Harmonic DenseNet), a derivative of DenseNet designed to cut memory traffic while keeping DenseNet’s feature reuse. The aim was to improve stroke segmentation accuracy over a standard UNet baseline, which was implemented and benchmarked alongside it.
Stack
Python · PyTorch