← Back

RESULTS

Everything below was measured on 100 patients the models never saw during training. Dice score shows how well the prediction overlaps the doctors' outline (1.0 = perfect).

0.885Mean Dice with all four MRI scans

0.015Active tumor Dice when T1ce is missing

0.434Whole tumor Dice when FLAIR is missing

0.391Active tumor without T1ce, after dropout training

1 · Normal model, one scan missing

Scans	Missing	Whole	Core	Active
All four	–	0.915	0.880	0.860
T1ce, T1, T2	FLAIR	0.434	0.571	0.581
FLAIR, T1, T2	T1ce	0.870	0.199	0.015
FLAIR, T1ce, T2	T1	0.898	0.856	0.836
FLAIR, T1ce, T1	T2	0.885	0.847	0.853

T1 and T2 can go missing with almost no damage. Without FLAIR the outline of the whole tumor breaks. Without T1ce the active part of the tumor almost disappears.

2 · Does the model notice?

Each scan was predicted 8 times with small flips. If the answers disagree, the model is unsure.

Condition	Dice	Uncertainty	Change
All four scans	0.885	0.0172	–
No FLAIR	0.545	0.0515	3× higher
No T1ce	0.363	0.0196	barely moves

Without FLAIR the model gets unsure, so it warns us. Without T1ce it stays confident while failing. That is the silent failure. (Spearman r = −0.115, p = 0.047.)

3 · The fix: dropout training

A second model was trained with random scans switched off, so it learns not to depend on just one.

Model	Missing	Whole	Core	Active
Normal	–	0.915	0.880	0.860
Dropout	–	0.911	0.837	0.855
Normal	FLAIR	0.434	0.571	0.581
Dropout	FLAIR	0.831	0.805	0.828
Normal	T1ce	0.870	0.199	0.015
Dropout	T1ce	0.892	0.563	0.391

Big gains when a scan is missing, and only a small cost when all scans are there. Still, 0.391 is far from 0.860: contrast enhancement is simply not visible without T1ce.

4 · Grad-CAM (a limitation)

Share of the model's attention that lands inside the real tumor, averaged over 100 patients.

Model	Scans	Whole	Core	Active
Normal	all four	36%	2%	5%
Normal	no T1ce	19%	3%	3%
Dropout	all four	13%	7%	4%
Dropout	no T1ce	18%	9%	4%

For the small regions (core and active) only a few percent of the attention lands inside the tumor, so Grad-CAM cannot explain them reliably. The conclusions here rest on Dice and uncertainty.

Data: BraTS 2021 (public, anonymized). 199 training, 49 validation, 100 test patients. Normal model trained 70 epochs, dropout model 100 epochs. Research demo, not for medical use.