Abstract
Accurate segmentation of lung tumours in computed tomography (CT) images is a key task in computer-aided diagnosis and treatment planning. This paper presents a comparative analysis of four autoencoder architectures—Vanilla Autoencoder, Variational Autoencoder (VAE), U-Net Autoencoder, and Attention-VAE—for the segmentation of lung cancer lesions using the Task03_Lung dataset from the Medical Segmentation Decathlon. Quantitative evaluation based on Dice Similarity Coefficient, Intersection over Union, Hausdorff Distance, Precision, Recall, and Volume Similarity reveals the superiority of the Attention-VAE model due to its ability to focus on semantically relevant features through attention mechanisms. The study also highlights the clinical interpretability and computational efficiency of autoencoder-based segmentation methods. The findings support the suitability of probabilistic and attention-guided autoencoding for high-precision medical image analysis.