Skip to content
Imtiaz Hossain

playground / computer vision / medical imaging

Brain Tumor Segmentation & Classification

U-Net segments the tumor, DenseNet-121 names it. Page through real inference on unseen MRI slices.

Full case study →Report (PDF) ↗
test dice
88.22%
U-Net segmentation
mIoU
79.74%
U-Net
pixel acc
99.61%
U-Net
classification
97.50%
DenseNet-121
dataset
BRISC 2025
/ 6,000 T1 MRI slices, 4 classes
test set
860
/ unseen samples
hardware
RTX 3070
/ 8GB, PyTorch 2.0 + CUDA 11.8
inference
~45 ms
/ per slice

interactive / live in your browser

Prediction overlay
Input MRI
Input MRIPrediction overlay

iThese are authentic saved outputs from the trained U-Net and DenseNet-121. Segmentation IoU is shown for test-set slices that carry ground-truth masks.

classifier output

Glioma99.9%
seg IoU
81.37%
ground truth
Glioma

Test slice with ground truth; strong overlap.

curated examples

the pipeline

From raw data to a verifiable result

  1. 01 / dataset

    BRISC 2025 MRI

    T1-weighted brain MRI at 256x256, spanning glioma, meningioma, pituitary, and no-tumor. 3,933 valid image-mask pairs drive the segmentation task.

    5,000 train / 1,000 test256x256 grayscale4 tumor categories
  2. 02 / augmentation

    Aggressive augmentation

    Horizontal and vertical flips, rotation up to 15 degrees, affine transforms, brightness/contrast jitter, and Gaussian noise expand the modest dataset without distorting anatomy.

  3. 03 / architecture

    U-Net + DenseNet-121

    A five-level U-Net with skip connections handles segmentation under a combined Dice-BCE loss. A separate DenseNet-121 classifies the slice. Separate models beat a shared encoder here.

    U-Net: dice-bce lossDenseNet: cross-entropytrained separately
  4. 04 / training

    100 epochs, early stopping

    Adam at 1e-4, batch size 16, up to 100 epochs with patience 15. A 20-run grid confirmed Adam at 5e-5 as optimal across optimizers.

  5. 05 / evaluation

    860 unseen samples

    U-Net won all four segmentation metrics against Attention U-Net. Test performance exceeded validation by ~5 points, evidence of robust training rather than overfitting.

  6. 06 / inference

    Original to prediction

    Below: real inference panels. Drag the slider to wipe between the input MRI and the model's predicted tumor overlay, and read the classifier's call and confidence.

evaluation artifacts

next experiment

AI-Generated Text Detection