chn123/bimcv-analysis

BIMCV-R 500-Series Multi-Model Analysis & Explainability Dataset

This repository contains end-to-end multi-model AI evaluation, lung cancer risk modeling, explainability maps (Grad-CAM), and automated radiology report generation for 500 randomly sampled physician-labeled chest CT series from the cyd0806/BIMCV-R dataset.

The cohort was evaluated across three state-of-the-art chest CT deep learning models:

1. Sybil (5-Seed Ensemble): 1-to-6 year lung cancer risk probability estimation and 3D attention/Grad-CAM explainability maps.
2. Pillar (Pillar0-Sybil-1.5 / YalaLab): 1-to-6 year lung cancer risk prediction using deep multi-window volumetric representations (11 CT windows).
3. Astra (Qwen2.5-VL-7B-Instruct + Merlin): Multimodal 3D vision-language model generating structured chest CT radiology reports (Findings and Impression).

---

📁 Repository Structure

├── BIMCV_Sybil_Pillar_Astra_500_Results.xlsx  # Master formatted Excel workbook (3 sheets, 500 cases)
├── bimcv_500_selected_metadata.csv           # Clinical metadata, demographics, and 95 physician pathology labels
├── BIMCV_FN_GRADS/                           # 3D and native Grad-CAM attention matrices (476 series, 16.3 GB)
│   ├── sub-S03406_ses-E77057_run-3_bp-chest_ct_gradcam.npz
│   ├── sub-S03657_ses-E76769_run-1_bp-chest_ct_gradcam.npz
│   └── ... (476 compressed .npz files)
├── astra_generated_reports/                  # 500 full-text structured radiology reports generated by Astra
│   ├── sub-S03406_ses-E77057_run-3_bp-chest_ct_astra_report.txt
│   ├── sub-S03657_ses-E76769_run-1_bp-chest_ct_astra_report.txt
│   └── ... (500 .txt files)
├── results_json/                             # Direct model prediction JSONs
│   ├── sybil_500_scores.json                 # Sybil 1-to-6 year probability scores & runtimes
│   ├── pillar_500_scores.json                # Pillar 1-to-6 year probability scores & runtimes
│   └── astra_500_reports.json                # Astra structured reports & runtimes
└── BIMCV_Sybil_Pillar_Astra_Results.xlsx      # Earlier 10-series prototype workbook

---

📊 Master Excel Workbook (BIMCV_Sybil_Pillar_Astra_500_Results.xlsx)

The primary deliverable is a comprehensive, publication-ready Excel spreadsheet containing 500 patient series across three dedicated worksheets:

Sheet 1: Skorlar ve Raporlar (Scores & Clinical Reports)

All key information is placed side-by-side on the exact same row:

• Identifier Data: Row Number, Patient ID (PatientID), Report ID (ReportID), CT Series Filename (ct_path).
• Physician Ground-Truth Labels: Summary string of active findings verified by radiologists.
• Sybil Risk Scores: 1st, 2nd, 3rd, 4th, 5th, and 6th-year estimated lung cancer risk probabilities.
• Pillar Risk Scores: 1st, 2nd, 3rd, 4th, 5th, and 6th-year estimated lung cancer risk probabilities.
• Original Spanish Clinical Report: Hospital ground-truth report (BIMCV_meta.csv).
• Original English Translated Report: Ground-truth report translated into English (Report_en).
• Astra AI Generated Report: Full structured report automatically written by Astra (Qwen2.5-VL-7B).
• Execution Runtimes: Sybil, Pillar, and Astra execution durations in seconds.

Sheet 2: Radyoloji Rapor Kıyaslaması (Full-Text Radiology Report Comparison)

• Side-by-side comparative layout containing Patient ID, Series Name, Physician Findings, Original Spanish Report, English Translated Report, and Astra AI Generated Radiology Report.

Sheet 3: Hekim Etiket Matrisi (95-Pathology Binary Matrix)

• Complete one-hot/binary ground-truth matrix for all 95 clinical findings (COVID-19, adenopathy, consolidation, atelectasis, nodule, emphysema, pleural effusion, etc.) annotated by board-certified radiologists.

---

🧠 Explainability Maps: BIMCV_FN_GRADS

The BIMCV_FN_GRADS/ directory contains volumetric Grad-CAM attention matrices for all cases where the Sybil 1-year cancer risk score is below 0.2 (low-risk / potential false-negative screening cases):

• Coverage: Exactly 476 out of 500 series (95.2% of the screening cohort).
• Total Size: 16.3 GB (compressed using NumPy .npz format).
• File Contents:
  • gradcam_native: Shape (25, 16, 16) float32. The exact latent feature-grid attention map computed by the Sybil ensemble (25 depth bins × 16×16 spatial feature grid).
  • gradcam_volume: Shape (N, 512, 512) float16. 3D trilinearly interpolated attention map aligned directly with the original CT slice volume grid ($N$ slices × 512 × 512).
  • image_attention_1: Shape (5, 1, 25, 256) float16. Raw multi-head attention over the 25 spatial slabs across the 5 ensemble seeds.
  • volume_attention_1: Shape (5, 1, 25) float16. Volume attention weights across the 25 depth slabs.
  • scores: float32 array of the 1-to-6 year predicted cancer risk probabilities.

How to Load and Visualize Grad-CAM in Python:

import numpy as np
import matplotlib.pyplot as plt

# Load a Grad-CAM file
grad_file = "BIMCV_FN_GRADS/sub-S04398_ses-E08742_run-2_bp-chest_ct_gradcam.npz"
data = np.load(grad_file)

print("Series:", data["series_name"])
print("6-Year Risk Scores:", data["scores"])
print("Native Grad-CAM Shape:", data["gradcam_native"].shape)    # (25, 16, 16)
print("Volume Grad-CAM Shape:", data["gradcam_volume"].shape)    # (N, 512, 512)

# Visualize maximum intensity projection (MIP) of attention
vol_attention = data["gradcam_volume"]
axial_mip = vol_attention.max(axis=0)

plt.figure(figsize=(6, 6))
plt.imshow(axial_mip, cmap="jet")
plt.title(f"Grad-CAM Axial MIP: {data['series_name']}\nYear 1 Risk: {data['scores'][0]:.4f}")
plt.colorbar(label="Sybil Attention Intensity")
plt.axis("off")
plt.show()

---

⚡ Multi-GPU Parallel Inference Architecture

To process 500 full 3D chest CT scans (~41 GB) efficiently, the pipeline was distributed across 3 dedicated NVIDIA A100-SXM4-40GB GPUs:

Model	Framework / Architecture	GPUs Used	Distribution Strategy	Total Duration (500 Scans)
Download	HTTP Range / RemoteZipRaw	8 CPU Workers	Concurrent chunk extraction across 40 zip archives	16.5 mins (41.3 MB/s)
Pillar	Pillar0-Sybil-1.5 (3 Seeds)	GPU 0, 1, 3	3 parallel workers (166 series/GPU)	6.2 mins (~1.5s/scan)
Sybil + Grad-CAM	Sybil Ensemble (5 Models)	GPU 0, 1, 3	3 parallel workers (166 series/GPU)	43.9 mins (~10s/scan)
Astra	Qwen2.5-VL-7B + Merlin	GPU 0, 1, 3	3 parallel workers (166 series/GPU)	79.2 mins (~28s/scan)
Excel & HF Push	OpenPyXL + HfApi	Multi-threaded	Automated workbook styling & artifact upload	3.4 mins
Total Pipeline	End-to-End	3 x A100	Fully Automated	133.2 mins (~2.2 hours)

---

📖 Citation & References

@article{mikhael2023sybil,
  title={Sybil: A Validated Deep Learning Model to Predict Future Lung Cancer Risk From a Single Low-Dose Chest Computed Tomography},
  author={Mikhael, Peter G and Wohlwend, Jeremy and Yala, Adam and others},
  journal={Journal of Clinical Oncology},
  volume={41},
  number={12},
  pages={2191--2200},
  year={2023}
}

@misc{agrawal2025pillar0,
  title         = {Pillar-0: A New Frontier for Radiology Foundation Models}, 
  author        = {Kumar Krishna Agrawal and Longchao Liu and Long Lian and Michael Nercessian and Natalia Harguindeguy and Yufu Wu and Peter Mikhael and Gigin Lin and Lecia V. Sequist and Florian Fintelmann and Trevor Darrell and Yutong Bai and Maggie Chung and Adam Yala},
  year          = {2025},
  eprint        = {2511.17803},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2511.17803}
},
  author={Yala, Adam and others},
  year={2024}
}

@misc{wang2026astrageneralizablereportgeneration,
  title         = {Astra: a generalizable report generation foundation model for 3D computed tomography}, 
  author        = {Zhuhao Wang and Fang Chen and Chaohui Yu and Zihan Li and Yuchao Zheng and Jing Wang and Xuan Yang and Jia Guo and Zhenlu Yang and Xingju Zheng and Yihua Sun and Haojie Han and Xiaoxiao Qin and Zhan Feng and Wenbo Xiao and Chao Zhu and Yuehua Li and Shipeng Zhang and Hao Luo and Yunsong Peng and Fan Wang and Hongen Liao},
  year          = {2026},
  eprint        = {2605.31437},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2605.31437}
},
  year={2024}
}

@article{delavaya2020bimcv,
  title   = {BIMCV-COVID19+: A large annotated dataset of RX and CT images from COVID-19 patients},
  author  = {de la Iglesia Vay{'a}, Mar{'\i}a and Saborit, Jos{'e} Mar{'\i}a and Montell, Jos{'e} Antonio and Pertusa, Antonio and Bustos, Aurelia and Cazorla, Miguel and Galant, Joaquin and Barber, Xavier and Orozco-Beltr{'a}n, Domingo and Garc{'\i}a-G{'o}mez, Juan M and Salinas, Joaqu{'\i}n Manuel},
  journal = {Medical Image Analysis},
  volume  = {69},
  pages   = {101979},
  year    = {2021}
}