How to Train a Tumor Segmentation Model on Ultrasound Data with MONAI
SMRTR summary
A MONAI-based tutorial walks through training a tumor segmentation model on the BUS-BRA breast ultrasound dataset — 1,875 images from 1,064 patients — by profiling the data first. A 10:1 background-to-tumor pixel imbalance drove the choice of DiceCELoss, while patient-grouped splits prevented data leakage, ultimately achieving a test Dice score of 0.864.
SMRTR provides this summary for quick context. The original article belongs to Daily.dev.
Read the original article