Deep Learning-Based Segmentation of Lymphangioleiomyomatosis (LAM) in Computed Tomography Scans Using V-Net

• Published in: International Conference on Inventive Computation Technologies (Online) pp. 1103 - 1109
• Main Authors: S, Sithika Seema, R, Sumathy
• Format: Conference Proceeding
• Language: English
• Published: IEEE 23.04.2025
• Subjects: Accuracy; Computed tomography; Computed Tomography (CT) Segmentation; Convolutional neural networks; Convolutional Neural Networks (CNNs); Deep learning; Image segmentation; Lymphangioleiomyomatosis (LAM); Medical diagnostic imaging; Medical Image Analysis; Sensitivity and specificity; Solid modeling; Three-dimensional displays; V-Net; Visualization
• ISSN: 2767-7788
• DOI: 10.1109/ICICT64420.2025.11004924

This Lymphangioleiomyomatosis is a rare lung disease which is characterized by abnormal proliferation of a smooth muscle - like cells and its leads to cystic lung destruction. Early and precise detection is pivotal for immediate assistance and functional medical management. Traditional diagnostic methods, such as biopsy and high-resolution computed tomography (HRCT), can be protracted and subject to inter - observer variability. The rapid progress in deep learning, especially Convolutional Neural Networks (CNNs), have improved precision in analyzing medical images. The above particular study plans to propose a V-Net based deep learning approach to the automatic detection and partition of LAM in CT scans. V - Net, a fully convolutional neural grid designed for volumetric clinical photograph partition, is trained on an annotated LAM CT scan dataset. The model is optimized using Dice loss to enhance segmentation accuracy. Performance evaluation is conducted using key metrics, including Dice Similarity Coefficient (DSC), Intersection over Union (IoU), sensitivity, and specificity. Experimental results demonstrate that V -Net outperforms traditional segmentation techniques in accurately identifying LAM-affected regions. By providing an automated, reliable segmentation tool, the proposed method reduces human error and improves diagnostic efficiency for radiologists. Hereafter work will pivot on enlarge the dataset, and conducting clinical validation to enhance robustness and generalizability.