TextBraTS-arch3: Lightweight 2.5D Text-Guided Brain Tumor Segmentation

A lightweight 2.5D convolutional network for brain tumor segmentation + radiology report generation on the TextBraTS/BraTS2020 dataset.

Results (93-patient official test set)

Metric	Value
Avg Dice	83.8%
Avg HD95	3.41 mm
ET Dice	79.7
WT Dice	89.5
TC Dice	82.3

Beats TextCSP SOTA (4.81mm) and TextBraTS (5.13mm) on HD95 at 10.4M parameters.

Repo structure

• segmentation/best_avg.pt — segmentation model checkpoint (epoch 64, val-selected)
• t5/model.safetensors + tokenizer — fine-tuned T5-small report generation head
• t5/img_proj.pt — image-conditioned projection weights for T5

Architecture

• Input: 2.5D — 4 MRI modalities × 3 adjacent axial slices = 12 channels at 128×128
• Encoder: 4-stage ResNet (32→64→128→256 ch)
• Text: Offline RadBERT embeddings (zero forward-pass text cost)
• Decoder: Attention-gated skips + soft cascade WT→TC→ET
• Report gen: ImageConditionedT5-small (8 image-prefix soft tokens)

Dataset

TextBraTS / BraTS2020 — official split: 220 train / 56 val / 93 test