Evaluating a Standard 3D U-Net for Lung Tumour Segmentation on the Medical Segmentation Decathlon Task06 Dataset

Lung tumour segmentation from computed tomography (CT) images is an important step in computer-aided diagnosis because it supports radiotherapy planning, treatment monitoring, and clinical decision-making. Over the past decade, numerous deep learning models, including attention-based and transformer-based architectures, have been proposed to improve segmentation accuracy. Despite these advances, comparisons between methods are often difficult because many studies introduce multiple architectural modifications without providing a transparent and reproducible baseline. In this work, we evaluate a MONAI implementation of a residual three-dimensional U-Net (3D U-Net) on the Medical Segmentation Decathlon (MSD) Task06 Lung dataset using three-fold cross-validation. The model achieved a mean Dice similarity coefficient of 0.2876 ± 0.0684, with performance particularly affected by severe foreground-background class imbalance and the presence of very small tumours. Although the baseline does not match the performance of more complex state-of-the-art models, it provides a fully reproducible reference implementation and highlights the challenges associated with lung tumour segmentation.