Bayesian automatic screening of pneumonia and lung lesions localization from CT scans

Author: Álvaro Moure Prado
Correspondence: ignacio.godino@upm.es
License: CC-BY-NC 4.0 International
bioRxiv Preprint: doi:10.1101/2025.04.08.647710 IEEE Access: 10.1109/ACCESS.2025.3607282

Overview

This repository accompanies the paper:

Bayesian automatic screening of pneumonia and lung lesions localization from CT scans. A combined method toward a more user-centred and explainable approach
Álvaro Moure Prado et al., IEEE Access, October, 2025

Abstract

While semantic segmentation allows precise lesion localization, bounding box-based object detection is considered more effective for highlighting target regions without replacing clinical expertise—reducing attentional and automation biases. This work proposes a two-stage approach for more explainable detection of pneumonia lesions in lung CT scans:

• Stage 1: Bayesian uncertainty-driven screening classifies each CT slice for disease presence.
• Stage 2: Lesion localization is applied to screened positive images using object detection architectures.

Key contributions:

• Explainability: Provides confidence measures for both predictions and regions of interest.
• Expert-centric: Supports, but does not replace, clinical judgement.
• Methodological innovation: Introduces a fusion strategy to merge overlapping bounding boxes, improving localization of scattered lesions.

Experiments were conducted on ~90,000 CT images from public COVID-19, bacterial, fungal, viral pneumonia datasets and controls.

---

Table of Contents

• Background
• Methods
• Datasets
• Installation
• Quick Start: Training & Inference
• Results
• Citation
• License
• Contact

---

Background

CT imaging provides high sensitivity for pneumonia diagnosis, especially COVID-19, enabling accurate localization of lung lesions. However, manual analysis is time-consuming and subjective. AI-powered systems can assist radiologists by automating screening and lesion localization, provided they communicate model uncertainty and do not induce clinical over-reliance.

This repo implements and benchmarks:

• Bayesian screening (DenseNet121 with uncertainty quantification)
• Object detection-based localization (YOLOv8, Cascade R-CNN, RetinaNet)
• Bounding box fusion to handle overlapping and scattered lesion regions

---

Methods

• Screening:
  Bayesian DenseNet121 classifies CT slices, estimating predictive uncertainty via Monte Carlo dropout (BayesianTorch).
• Localization:
  Positive slices are processed by YOLOv8, Cascade R-CNN, or RetinaNet to detect lesions as bounding boxes.
• BB Fusion:
  Semantic segmentation masks are converted to bounding boxes using connected components, Non-Maximum Suppression, and HDBSCAN clustering.

Pipeline:

1. Pre-process images (CLAHE, histogram equalization)
2. Screen slices for disease (COVID-19, CAP, HC)
3. Apply object detector to positive cases
4. Post-process bounding boxes for robust lesion localization

---

Datasets

Eight public datasets were curated and harmonized with segmentation mask-to-bounding box conversion.

• COVID-19, CAP, HC, and control subjects
• Details and code for annotation conversion provided in the repo.

See the paper for dataset specifics.

---

Installation

1. Clone this repository:

   git clone https://github.com/BYO-UPM/CT-COVID.git
   cd <repo-name>

2. Install dependencies:
   (Recommended: use a Python 3.9+ virtual environment)

   pip install -r requirements.txt

3. Download datasets:

   • Depending on the time of reading, some data sources might have changes its access policies.

---

Quick Start: Training & Inference

1. Dataset Creation

For dataset creation please check: data_processing/gen_classification_dataset.ipynb

and

data_processing/gen_object_detection_dataset.ipynb

Original CT Images need to be downloaded from the original sources independently.

2. Screening (Bayesian DenseNet121)

Train screening module:

python main_lightning.py **kwargs

• Configurable for deterministic or Bayesian mode.
• Outputs accuracy, ROC curves, and uncertainty plots.

3. Lesion Localization (YOLOv8 / Cascade R-CNN / RetinaNet)

Train lesion localization module (MMDET-based):

python src/det/models/mmdet/train.py

Train lesion localization module (YOLO-based):

src/det/models/yolo/train_yolo.py

---

Citation

If you use this work, please cite:

@article{moureprado2025bayesian,
  author    = {Álvaro Moure Prado, Alejandro Guerrero-López, Julián D. Arias-Londoño, and Juan I. Godino-Llorente},
  title     = {Bayesian automatic screening of pneumonia and lung lesions localization from CT scans. A combined method toward a more user-centred and explainable approach},
  journal   = {IEEE Access},
  year      = {2025},
  doi       = {10.1109/ACCESS.2025.3607282}
}

---

License

This repository is licensed under CC-BY-NC 4.0 International.

---

Contact

For questions and collaborations, please contact:
ignacio.godino@upm.es

---

Keywords: Pneumonia | COVID-19 | Lung Lesion Localization | CT Scan | Explainable AI | Bayesian Deep Learning | Object Detection | Decision Support System

---

Let me know if you want specific code snippets, a more detailed usage section, or additional badges!