brain-tumor-deeplabv3plus

Brain Tumor Segmentation — DeepLabv3+ (ResNet-34 + ASPP + WBCE)

An implementation of the architecture and training recipe described in Soomro et al., "Boosting Brain Tumor Detection Accuracy in MRI Using Transfer Learning and Fine-Tuned DeepLabv3+" (IEEE Open Journal of the Computer Society, 2026), trained on the CE-MRI brain tumor dataset (Cheng et al., 233 patients, 3,064 T1-weighted contrast-enhanced axial slices: glioma, meningioma, pituitary tumors).

This is a personal/academic reproduction project — not the original authors' model, not clinically validated, and not for medical use. See Honest results below before using this for anything beyond learning/experimentation.

Model architecture

• ResNet-34 encoder, initialized from ImageNet-pretrained weights
• Atrous Spatial Pyramid Pooling (ASPP), rates 12/16/18 + global image pooling
• Lightweight decoder fusing low-level and ASPP features (DeepLabv3+ design)
• Binary output: tumor vs. background, per pixel
• Trained with Weighted Binary Cross-Entropy (WBCE) loss, with per-slice adaptive class-imbalance weighting

Training setup

• 100 epochs, Adam optimizer (β1=0.9, β2=0.999, weight decay 1e-5), initial LR 1e-4 with ReduceLROnPlateau
• Patient-wise 70/15/15 train/val/test split (163/35/35 patients) — no patient's slices appear in more than one split, avoiding data leakage
• Image size 512×512, batch size 8
• Trained on a single Kaggle GPU session (T4×2)

Full training/evaluation code: [link to your GitHub repo here]

Honest results vs. the paper

The paper reports ~98% DSC and ~99.3% sensitivity. This reproduction falls meaningfully short of that on the two metrics that matter most for segmentation quality (DSC, sensitivity), while matching closely on accuracy/specificity — which is expected, since ~98% of pixels in these images are background, so accuracy/specificity are dominated by the easy majority class rather than tumor-finding ability.

Per-tumor-type DSC: glioma 70.2%, meningioma 85.5%, pituitary 74.6% — the same relative ordering (meningioma easiest, glioma hardest) as the paper reports, which is a useful internal consistency check even though absolute numbers differ.

Why the gap likely exists (in rough order of suspected impact):

1. The paper doesn't fully specify some architectural details (exact decoder channel widths, output stride choice, precise augmentation policy) — this implementation fills those gaps with standard DeepLabv3+ conventions, which may not match the authors' exact setup.
2. Training was extended from 40 to 100 epochs and DSC barely moved (74.3% → 75.6%), suggesting the model has converged for this configuration — the remaining gap is not simply "needs more training."
3. Possible differences in preprocessing intensity (the paper mentions CLAHE contrast enhancement; exact parameters aren't given) or data augmentation strength.

This gap is reported transparently rather than hidden — reproducing a paper's exact numbers without the authors' full implementation details is a known hard problem in ML research, and getting a rigorous, honest measurement of how far off a reproduction is is itself the useful skill being demonstrated here.

Intended use

• Educational / portfolio demonstration of a segmentation pipeline (transfer learning, class-imbalance-aware loss, patient-wise evaluation discipline)
• Starting point for further experimentation (e.g. testing architectural variants, better augmentation, longer training with a different LR schedule)

Out of scope / not suitable for

• Any clinical, diagnostic, or medical decision-making use
• Deployment without independent validation on a properly consented, IRB-approved clinical dataset
• Use as a certified medical device (it is not one, and has not been evaluated as one)

How to use

import torch
from model import DeepLabV3Plus  # from this repo's model.py

model = DeepLabV3Plus(num_classes=1, use_imagenet_init=False)
state = torch.load("pytorch_model.pt", map_location="cpu", weights_only=False)
model.load_state_dict(state)
model.eval()

# image: torch.Tensor, shape (1, 3, 512, 512), ImageNet-normalized
with torch.no_grad():
    logits = model(image)
    mask = (torch.sigmoid(logits) > 0.5).float()

See the GitHub repo for the full dataset.py preprocessing pipeline (CLAHE + ImageNet normalization) needed to prepare inputs correctly.

Dataset

Trained on the CE-MRI brain tumor dataset (Cheng et al., 2017, "Enhanced performance of brain tumor classification via tumor region augmentation and partition," PLoS ONE). This repository does not redistribute the dataset — see the original publication for access.

Citation

If referencing the original paper this reproduces:

Soomro et al., "Boosting Brain Tumor Detection Accuracy in MRI Using
Transfer Learning and Fine-Tuned DeepLabv3+," IEEE Open Journal of the
Computer Society, vol. 7, 2026.

This repository is an independent reproduction and is not affiliated with the original authors.