yssvyss/HEADLINE-MSIIP

MSIIP HEADLINE 2026 model weights

Research weights for Entity-Aware Head CT Report Generation via Classification-Derived Label Priors.

The repository contains a Qwen3.5-4B report generator, eight auxiliary multi-label classifier checkpoints, and the visual backbone files needed by the inference pipeline. The generator language module was fully fine-tuned while its vision encoder and MLP projector were frozen. The generator shards were verified against the frozen-vision checkpoint-1287 release candidate.

The model files are designed for the MSIIP HEADLINE v5 inference container. Download the repository into the directory mounted at /opt/ml/model; vlm/ and multilabel/ should be directly under that directory. Inputs follow the HEADLINE soft-tissue CT and geometry socket layout.

The models are intended for research and benchmarking. They are not validated for clinical use. The model may miss findings, produce unsupported findings, or misinterpret candidate labels. No clinical images or reports are included.

The release includes modified or extracted model weights from the Qwen family. See the upstream Qwen3.5-4B and Qwen3-VL-4B-Instruct model cards and license notices.