Results

Every number, every curve, every prediction

Test-set metrics for the U-Net baseline, the structural comparison across all five models, 35 epochs of training telemetry for each run, and the prediction galleries they produced. All values are read directly from the notebook cell outputs.

Headline metrics Training curves Prediction gallery

Comparison

Structural agreement with expert masks

All five models scored on tumour-bearing validation slices, comparing the argmax label map of the prediction against the expert annotation.

PSNR (dB)

higher is better

Model A · U-Net25.38
Model D · 4×425.96
Model D + CLAHE24.76
Model B · 3×3 s224.64
Model C · 3×3 s121.62

Bars scaled across the observed range 21.5–26.0 dB to make the differences legible; the axis does not start at zero.

SSIM

higher is better

Model A · U-Net0.9577
Model D · 4×40.9583
Model D + CLAHE0.9582
Model B · 3×3 s20.9510
Model C · 3×3 s10.9341

Scaled across 0.933–0.959. The top three sit within 0.0001 of each other — well inside run-to-run noise.

Full comparison. Segmentation loss is the generator’s cross-entropy term at the final epoch; training time is wall-clock on a Kaggle GPU session. Model C holds the best seg loss and the worst SSIM — the PatchGAN Paradox.
Model	Discriminator	Params (D)	Seg loss	PSNR	SSIM	Train	Final D loss
Model A — U-Net	none — pixel-wise loss	—	0.0177	25.3817	0.9577	88 min	—
Model D — 4×4, s2	4×4 kernel, stride 2	727,489	0.0198	25.9556	0.9583	137 min	1.1529
Model D + CLAHE	4×4 kernel, stride 2	727,489	0.0194	24.7633	0.9582	147 min	1.0928
Model B — 3×3, s2	16×16 patch grid	1,560,897	0.0378	24.6395	0.9510	148 min	1.0109
Model C — 3×3, s1	16×16 patch grid	1,559,937	0.0355	21.6154	0.9341	211 min	12.4787

Read these gaps as directional, not exact. The train/validation split was drawn without a fixed random seed, so each notebook evaluated on a different partition — 4,453 tumour slices for Model A, 4,623 for Model B, 4,524 for Model C. Sub-decibel differences sit inside that variance. The 3 dB collapse of Model C does not, and neither does the artefact / no-artefact distinction. Where the accompanying paper quotes different figures for Model D, the notebook-measured value is used throughout this site. Full limitations →

Model A · baseline U-Net

Model A on the held-out test set — 45 patients, 4,500 slices

Accuracy

99.27%

Precision

99.34%

Sensitivity

99.11%

Specificity

99.78%

Per-class Dice coefficient on the test set.
Tumour sub-region	Dice
GD-enhancing tumour	0.7084
Peritumoral edema	0.6995
Necrotic / non-enhancing core	0.5687
Mean over all four classes	0.6124

Reading the numbers honestly

Accuracy near 99 % sounds decisive and is almost meaningless here: predicting "background" for every voxel would already score ~98 %. Dice is the number that matters, and 0.57–0.71 per sub-region is a solid result for a compact 2D model trained for 35 epochs on downsampled input — well short of the 0.85+ that 3D ensembles reach on the BraTS leaderboard.

The ordering across sub-regions is itself a sanity check on the pipeline. The enhancing rim is brightest on T1ce, edema is largest and cleanest on FLAIR, and the necrotic core is small, irregular and defined by the absence of enhancement — hardest for all three reasons at once.

Mean IoU is omitted deliberately. The pipeline configures tf.keras.metrics.MeanIoU(num_classes=4) with one-hot targets and softmax probabilities instead of integer class indices, so it degenerates to a constant 0.5. No IoU figure from these runs is trustworthy. Why →

Training telemetry

35 epochs, five runs

Reconstructed from the printed epoch logs of each notebook and the baseline's training.log. Click a legend entry to isolate a run; hover the plot to read exact values.

The discriminator plot is the most informative chart in this study. Models B and D settle near 1.0 — the equilibrium value when a critic is guessing at chance, which is exactly where a healthy GAN lands. Model C instead climbs from 9.06 to 12.48: it is getting steadily worse at its own task while the generator's segmentation loss keeps falling. That divergence is the PatchGAN Paradox in telemetry form, and the mechanism behind Model C's 3 dB PSNR deficit.

Baseline learning rate drops are visible as the flattening after epoch ~24, when ReduceLROnPlateau cut it by 5× twice in succession down to the 10^-6 floor. The best checkpoint by validation loss was epoch 21 (val_loss 0.026537).

Prediction gallery

Side by side, on unseen scans

Every figure below is an unmodified output cell from the notebook that produced it. Click any image to open it full size.

Necrotic / core
Peritumoral edema
GD-enhancing tumour
Colour maps vary between notebooks — each figure carries its own scale

Model D — the Goldilocks Zone

Three rows: axial MRI slice, expert ground truth mask, and 4x4 cGAN prediction. The predicted core and edema regions closely match the annotation in shape and position. — **Model D · 4×4, stride 2.** Topology is right in all three cases; the model smooths away the small detached fragment visible in the middle row's ground truth — a direct consequence of the shape prior preferring one connected lesion.

Three rows in the same layout for the CLAHE-enhanced 4x4 cGAN variant. — **Model D + CLAHE.** Visually near-identical to the run above, which is consistent with the metrics: SSIM moved by 0.0001 while PSNR fell 1.2 dB — the perception–distortion trade-off.

Models B and C — the two failure modes

One row: MRI slice, ground truth mask and 3x3 PatchGAN cGAN prediction, with visibly rougher region boundaries. — **Model B · 3×3, stride 2.** Boundaries are sharper than Model D produces, but class assignment inside the lesion is noticeably rougher — the structural cost of the uneven 3 ÷ 2 tiling.

The same three-panel layout for the pooled 3x3 PatchGAN variant. — **Model C · 3×3, stride 1 + pool.** The PatchGAN Paradox in one frame: the lesion is located and locally crisp, but internal class structure fragments.

Model A — per-class probability maps

Six panels per patient: the FLAIR slice, the expert overlay, all predicted classes together, then necrotic core, edema and enhancing tumour separately.

Six-panel U-Net prediction breakdown for a test patient with two distinct lesions. — Test patient with two separate lesions — both recovered.

Six-panel U-Net prediction breakdown for a second test patient. — A low-contrast case; the model still separates edema from core.

Six-panel U-Net prediction breakdown for a third test patient. — Third held-out patient.

Six-panel U-Net prediction breakdown for a fourth test patient. — Fourth held-out patient.

Six panels showing the original expert segmentation next to the predicted segmentation for all classes and then each class channel separately. — **Class-channel decomposition.** The expert mask, the combined prediction, then each of the four softmax channels on its own — the clearest view of what the network actually outputs before `argmax`.

As-run training plots

Three matplotlib plots from the notebook: training and validation accuracy, loss, and Dice coefficient over 35 epochs. — **Model A, as plotted in the notebook.** Accuracy, loss and global Dice, training against validation.

Matplotlib plot of generator, discriminator and segmentation loss for the 4x4 cGAN across 35 epochs. — **Model D, as plotted in the notebook.** Generator, discriminator and segmentation loss on shared axes.

Data exploration

What the pipeline sees before training starts

T1, T1ce, T2, FLAIR and the expert mask for one axial slice. — All four MRI sequences plus the expert mask for one slice. Only T1ce and FLAIR reach the network.

The same T1ce volume viewed in transverse, frontal and sagittal planes. — One volume in all three anatomical planes. The axial plane is used because it is natively square at 240×240.

Five panels isolating background, non-enhancing tumour, edema and enhancing tumour from the full mask. — The four label classes isolated from one mask — the nesting structure the discriminator is asked to learn.

Bar chart of patient counts: 250 training, 74 validation, 45 test. — The 250 / 74 / 45 patient split. Drawn without a fixed seed — the study's principal methodological weakness.