3D Lung Nodule Segmentation Using Difference over Union Combo Loss for UNet

Date of Award

5-1-2025

Degree Name

M.S. in Electrical Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Russell Hardie

Abstract

Lung cancer is the leading cause of cancer-related deaths. Which is why early and accurate detection of pulmonary nodules in computed tomography (CT) scans is necessary. Manual segmentation of nodules by radiologists is time consuming and prone to variability, fueling the need for automated solutions. This thesis introduces a novel Difference-over Union (DoU) Combo Loss function for 3D lung nodule segmentation using a 3D U-Net architecture, trained on the LIDC-IDRI dataset. The proposed loss function combines dice loss, binary cross-entropy, and boundary DoU loss to address region, pixel, and boundary level segmentation challenges, enhancing performance on imbalanced data. Quantitative results show that the DoU Combo Loss slightly outperformed the standard Combo Loss and other baselines, like the unified focal loss. Qualitative analysis showed reduced hallucination in complex CT scans with DoU Combo Loss, improving reliability. Although the quantitative gains are modest, the proposed method demonstrates potential for robust automated nodule segmentation, offering a simpler alternative to complex architectures and paving the way for enhanced clinical diagnostics.

Keywords

Electrical Engineering, Medical Imaging

Rights Statement

Copyright 2025, author.

Share

COinS