Moodski/bn4601a-stroke-models

BN4601A stroke MRI assistant: model weights

Research prototype, not a medical device and not for clinical use.

• unet/: 2.5D U-Net with a ResNet-34 encoder (ImageNet initialisation) that segments acute ischaemic stroke lesions.
  • Input: DWI, ADC and FLAIR, each slice with its neighbours, 1.5 mm in-plane, z-scored within the brain mask.
  • Output: lesion probability, thresholded at the value chosen on validation (summary.json).
  • Trained on ISLES 2022 and SOOP.
  • Dice 0.73 on the internal test set (246 scans) and 0.84 on a held-out hospital (Bern, 52 scans).
• adapter/: QLoRA adapter (rank 16) for Qwen2.5-VL-3B-Instruct, applied to the language model, the vision encoder and the merger.
  • Trained to answer questions grounded in structured findings and a cited knowledge base, and to perceive lesions on DWI | ADC | FLAIR images.
  • On held-out cases, the share of answers grounded in the findings rose from 8% (base model) to 98%, and out-of-scope questions were declined 100% of the time (vs 11% for the base model).

Demo: Moodski/bn4601a-stroke-assistant

Limitations

• Not evaluated on stroke-free scans.
• Weaker on post-treatment follow-up scans (external ISLES'24 test: Dice 0.66 mean, 0.75 median).
• No haemorrhage assessment.
• The DWI–FLAIR timing signal is a proxy that has not been validated against onset times.

Training data: ISLES 2022 (CC BY 4.0), SOOP (CC0). The base model's licence (Qwen research licence) applies to the adapter.