CAP 5516 Medical Image Computing2025

Volumetric tumor segmentation, and why the spread matters

A 3D U-Net trained with five fold cross validation on brain tumor MRI, where the spread of per case scores turned out to carry far more information than the average.

3D U-NetMONAIBraTSCross validationSLURM

0.6181

Whole tumor Dice

5 fold average

0.4472

Tumor core Dice

0.4524

Enhancing tumor

8.85 mm

Whole tumor HD95

The question

Tumor subregions occupy a tiny fraction of a brain volume, and the three targets are nested: enhancing tumor inside tumor core inside whole tumor. A model can post a respectable average while the smallest and most clinically important region goes unrecovered.

Approach

A 3D U-Net implemented in MONAI, trained with five fold cross validation on GPU nodes through SLURM, with augmentation covering rotations, flips and additive noise, and a cyclic learning rate schedule. I logged per case scores rather than only per fold means, which is what made the next part visible.

Four MRI modalities and the ground truth tumor mask — The four input modalities, FLAIR, T1w, T1gd and T2w, beside the ground truth mask. The tumor is legible in some channels and nearly invisible in others, which is the whole reason the input is multi modal.

Training loss curves across folds — Training loss per fold. It falls from about 0.77 to 0.36 by epoch 16, a 52 percent reduction, so the optimisation itself was healthy even where the validation Dice was not.

Outcome

Average best whole tumor Dice of 0.6181 across folds. Per case, whole tumor averaged 0.6340 while enhancing tumor averaged 0.4524, and individual cases ranged from 0.08 to 0.84. The distribution is bimodal rather than tight, which is precisely what a single mean averages away.

What it taught me

This is where I started reporting distributions instead of averages, and it is the habit that carried straight into my thesis work. A 0.63 mean that is really a mixture of confident successes and near misses calls for a different fix from one that is uniformly middling.

More output

Tumor segmentation prediction — Prediction against ground truth on a held out fold.

Multi modal MRI samples — Sample volumes across the four MRI modalities.

Dice across folds — Dice for enhancing tumor, tumor core and whole tumor across folds.

HD95 across folds — Hausdorff distance, which is far less forgiving than Dice.

Code and report BraTS

Next: What attention gates actually buy you