GRPO-tuned label-to-report assembler

Part of the master's thesis From Scan to Report: Detecting, Segmenting and Reporting Findings in Chest CT (code: https://github.com/Switchharsh/Switchharsh-From_Scan_to_Report_Master_Thesis; thesis: https://github.com/Switchharsh/Switchharsh-From_Scan_to_Report_Master_Thesis/tree/main/thesis).

The label assembler (scan2report-label-assembler) further trained with GRPO using a RadBERT-based CRG reward. Prompt: the bare label list Findings present: ... Findings absent: ... (no task sentence). Only the LoRA adapter is included (the checkpoint is a plain PEFT language model).

labels given CRG macro-F1
reference labels 0.9348 0.9628
FlexiCT attention classifier, t=0.60, raw text 0.5447 0.5558
same, with coverage post-processing 0.5461 0.5563
another patient's classifier labels 0.3507 -

GRPO raises the CRG gap between correct and mismatched labels (+0.0466 over the supervised assembler, Holm p=0.0003) but not the GREEN or CRIMSON gaps; it writes more findings per report (5.40 vs 4.66).

How it fits together

  • Input: the 8x8x8 FlexiCT token grid of a chest CT (flexict/tokens8, 864-dim per token) from CT-RATE.
  • projector.pt: MLP (864 -> 2560 -> 2560, LayerNorm) that maps the visual tokens into the LLM embedding space.
  • adapter_model.safetensors + adapter_config.json: LoRA (r=16, alpha=32) on Qwen/Qwen3.5-4B.
  • Loading and generation code (model wrapper, data, evaluation): https://github.com/Switchharsh/Switchharsh-From_Scan_to_Report_Master_Thesis (ctbench/rgen/model.py, evaluate.py). The adapter is not a stand-alone chat model: the projector and the FlexiCT tokens are required.
  • Task instruction used in every run: Write the findings and impression for this chest CT. Greedy decoding, at most 512 new tokens.

Evaluation protocol

1,000 patient-sampled CT-RATE validation cases; CRG (micro) and macro-F1 with the RadBERT label reader (suitch/radbert-english-ctrate-classifier, threshold 0.5). Control = another patient's scan tokens (mismatched). Numbers are from the thesis From Scan to Report: Detecting, Segmenting and Reporting Findings in Chest CT.

Limitations / licence

Research use only, not for clinical use. Trained on CT-RATE (CC BY-NC-SA 4.0), so the weights are released under the same non-commercial terms. Reports can contain false findings; see the thesis for the error analysis.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for suitch/scan2report-label-assembler-grpo

Finetuned
Qwen/Qwen3.5-4B
Adapter
(1,228)
this model

Collection including suitch/scan2report-label-assembler-grpo