MediSeg: Automated Analysis of Pulmonary Nodules with 3D Imaging

Abstract

The MediSeg project introduces an automated and interactive system for detecting, segmenting, and analyzing pulmonary nodules from 3D computed tomography (CT) scans using the LIDC-IDRI dataset. Developed in Python, this platform features a user-friendly GUI for patient selection, slice browsing, filter application, semi-automated segmentation, and generation of 2D–3D visualizations and comprehensive clinical PDF reports. Nodule segmentation is achieved through a classical region-growing algorithm seeded from expert-annotated ground truth, refined by 3×3 kernel morphological operations to improve boundary accuracy.
Unlike deep learning-based approaches like U-Net, which require extensive training and computational resources, MediSeg prioritizes interpretability and real-time processing, making it suitable for educational and low-resource settings.
In contrast to commercial CAD systems focused primarily on speed, MediSeg offers enhanced visualization by integrating vedo for 3D rendering, pyradiomics for radiomic feature extraction, and compatibility with ITK-Snap via NIfTI and CSV outputs., offering dynamic anatomical localization using a human-body overlay and multiple contrast/filter controls (e.g., Gaussian blur, Sobel, Laplacian, histogram equalization, Fourier-based filters).
The system incorporates a Random Forest classifier for initial candidate detection and a multi-nodule tracking algorithm across adjacent slices. Additional modules support real-time evaluation of segmentation performance using metrics like Dice, Jaccard, sensitivity, specificity, precision, F1-score, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR). Results are displayed interactively, enabling immediate interpretation and feedback.
Validation included 20 patients: 16 from the 50-patient training pool (e.g., 0001–0049) and 4 external cases (0052–0055), a subset of the 1018-patient LIDC-IDRI database, constrained by memory limitations (~100–200 MB per patient). Over 2000 images (~100/patient) yielded average metrics of 99.4% Dice, 99.0% Jaccard, 99.5% sensitivity, 100% specificity, 99.2% precision, 99.3% F1-score, and 100% accuracy. A smaller validation set of 10 patients showed a Dice score of 0.997 and perfect segmentation in several cases.
Some edge cases showed limitations (Dice = 0.72), prompting future upgrades with adaptive thresholds and deep learning. Complex cases (e.g., 0049, 0044) revealed over-segmentation or missed borders due to low contrast, irregular margins, or adjacent noise. These issues stem in part from a fixed 0.05 threshold, which may underperform in edge cases.
Though MediSeg is not yet used in classroom or clinical environments, it received feedback from Prof. Alberto Isaac Pérez Sanpablo. Although not yet tested with end-users, feedback from an academic expert confirmed the platform’s educational value. Its interactive GUI and dynamic visualization tools distinguish it from static academic systems, offering strong potential for biomedical imaging education.
The platform anonymizes all public data used and avoids any diagnostic or clinical claims. Future iterations will incorporate accessibility features and multi-user evaluation to promote equitable and ethical adoption.
The system has not yet been deployed in institutional settings not due to software limitations, but rather because broader testing and outreach are still pending. While current memory usage (~100–200 MB/patient) can limit model training or full-volume rendering on low-end machines, this is dependent on user hardware and does not affect the system’s general usability.