demo

# ReXGroundingCT Two-Stage Prototype Backup This dataset repository stores a backup of the local two-stage ReXGroundingCT prototype. ## Files - `rexgroundingct_project_lite.zip` - Core project archive. - Includes scripts, docs, manifests, evaluation outputs, visualizations, and selected project data. - Excludes large reusable model/data caches and intermediate Stage 2 epoch checkpoints. - `rexgroundingct_modelscope_models.zip` - ModelScope model cache used by the project. - Includes the Qwen embedding model files that were skipped by the lite project archive. - `rexgroundingct_2d_repo.zip` - Local ReXGroundingCT 2D data repository snapshot. - `rexgroundingct_stage2_checkpoints.zip` - Stage 2 STU-Net epoch checkpoint files from `outputs/stage2_stunet_full/epoch_*.pt`. ## Current Experiment The validated experiment is the 50-CT subset run under `subset50_fixed` output names. Key local outputs include: - `outputs/stage1_voxtell_baseline_subset50_fixed/` - `outputs/stage1_voxtell_proposals_subset50_fixed/proposal_manifest.jsonl` - `outputs/stage2_on_stage1_proposals_subset50_fixed/stage2_result_manifest.jsonl` - `outputs/stage2_on_stage1_proposals_clustered_subset50_fixed/` - `outputs/two_stage_subset50_fixed_summary.json` - `outputs/eval_two_stage_pipeline_subset50_fixed_full.json` - `outputs/visualizations/visualization_index.csv` - `outputs/visualizations/visualization_index.json` ## Notes Stage 1 uses VoxTell as a prompt-conditioned coarse localizer. Stage 1 masks are converted into connected-component ROI boxes, and Stage 2 uses STU-Net to segment each ROI. The current bottleneck is Stage 1 recall. Stage 2 improves results when the Stage 1 proposal covers the target, but cannot recover lesions outside the ROI.