soumickmj/GPReconResNet_BraTS2020T1ce_Axial

GPReconResNet — BraTS 2020 · T1CE (single contrast) · Axial

Architecture — GPReconResNet

GPReconResNet is a residual reconstruction network adapted for classification:

Hyperparameter	Value
Residual blocks	14
Starting feature maps	64
Up/down-sampling blocks	2
Activation	Leaky ReLU
Dropout (residual)	0.5
Upsampling	sinc interpolation
Batch normalisation	✗
3-D mode	✗ (2-D)

The reconstruction bottleneck forces the network to learn compact, semantically rich representations that are easy to interpret as class-discriminative maps.

Overview

This model is part of the GPModels family — a set of inherently-explainable convolutional networks for simultaneous brain-tumour classification and weakly-supervised segmentation from multi-contrast MRI. The models were trained and evaluated on the BraTS 2020 dataset.

⚠️ Single-contrast model — not used in the paper

This model was trained on T1CE contrast only (1 input channel) as an exploratory experiment. It is not part of the original publication: Weakly-supervised segmentation using inherently-explainable classification models and their application to brain tumour classification, which exclusively uses 4-contrast (T1 · T2 · T1CE · FLAIR) models.

• If you want a quick single-contrast experiment, this model may be useful.
• If you want to reproduce the paper results, or need the best performance, please use the 4-contrast counterpart: GPReconResNet — 4-contrast

• Input contrast: T1CE only (1 channel)
• Task: Multi-class brain-tumour classification (Healthy / LGG / HGG) and weakly-supervised segmentation
• Orientation: Axial 2-D slices (240 × 240 px)
• Output classes: 0 = Healthy, 1 = LGG (Low-grade glioma), 2 = HGG (High-grade glioma)
• Normalization: Per-image max normalization (divide by slice maximum)
• Preprint (open access): https://arxiv.org/abs/2206.05148
• Code & training details: https://github.com/soumickmj/GPModels

Model Inputs / Outputs

Property	Value
Input shape	(B, 1, 240, 240) — float32, max-normalized, T1CE only
Output — train mode	(B, 3) — raw logits
Output — eval mode	((B, 3), (B, 3, H, W)) — logits and spatial heatmap
Class order	[Healthy (0), LGG (1), HGG (2)]

Usage

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("soumickmj/GPReconResNet_BraTS2020T1ce_Axial", trust_remote_code=True)
model.eval()

# x: (B, 1, 240, 240) — T1CE slice, max-normalised
x = torch.randn(1, 1, 240, 240)
with torch.no_grad():
    logits, heatmap = model(x)     # eval mode: logits (B,3)  +  heatmap (B,3,H,W)

pred_class = logits.argmax(dim=1)  # 0=Healthy, 1=LGG, 2=HGG

Weakly-Supervised Segmentation (Heatmap Mode)

These models are inherently explainable: eval mode bypasses the global max-pooling (GMP) layer, exposing a full-resolution spatial activation map per class. No extra labels or re-training are required.

How it works:

• model.train() → GMP applied → returns logits (B, 3) — classification only
• model.eval() → GMP skipped → returns (logits, heatmap) where heatmap has shape (B, 3, H, W) — one spatial map per class

Heatmap channel order: 0 = Healthy, 1 = LGG, 2 = HGG. The whole-tumour map is obtained by combining channels 1 and 2.

with torch.no_grad():
    logits, heatmap = model(x)             # heatmap: (B, 3, H, W)

# Whole-tumour map (max over LGG + HGG channels)
wt_map  = heatmap[:, 1:, :, :].max(dim=1).values   # (B, H, W)

# Min-max normalise to [0, 1]
wt_flat = wt_map.view(wt_map.size(0), -1)
wt_min  = wt_flat.min(dim=1).values[:, None, None]
wt_max  = wt_flat.max(dim=1).values[:, None, None]
wt_norm = (wt_map - wt_min) / (wt_max - wt_min + 1e-8)

# Binary mask via threshold
binary_mask = (wt_norm > 0.5).float()               # (B, H, W)

For advanced post-processing (multi-Otsu, top-k binarisation, morphological clean-up, per-slice aggregation) see the project repository.

Training Details

Setting	Value
Trained by	Soumick Chatterjee (Pavan)
Dataset	BraTS 2020 — non-empty axial slices
Input contrast	T1CE only (1 channel)
Split	Stratified 75/25 train-test (seed 13); 5-fold CV (fold 0 reported)
Optimizer	Adam (lr = 1e-3, weight_decay = 5e-4)
Loss	Cross-entropy with class-balanced weights
Mixed precision	AMP
Max epochs	300

Citation

If you use this model, please cite:

@article{chatterjee2026weakly,
  title={Weakly-supervised segmentation using inherently-explainable classification models and their application to brain tumour classification},
  author={Chatterjee, Soumick and Yassin, Hadya and Dubost, Florian and N{\"u}rnberger, Andreas and Speck, Oliver},
  journal={Neurocomputing},
  pages={133460},
  year={2026},
  publisher={Elsevier}
}

License

MIT — see https://github.com/soumickmj/GPModels for full details.