Transforming liver tumor segmentation: A comprehensive meta-analysis of deep learning approaches.

Abstract

e13653 Background: Precise segmentation of hepatic neoplasms is essential for effective diagnosis and treatment planning. Manual segmentation of abdominal computed tomography (CT) images is time-intensive, subject to variability, and reliant on expert radiologists. Recent advancements in deep learning, particularly convolutional neural networks (CNNs), have demonstrated the potential to automate this process. However, challenges such as dataset heterogeneity, evaluation metrics, and clinical generalizability persist. This meta-analysis evaluates the efficacy of advanced neural network-based methodologies for hepatic neoplasm segmentation across standardized datasets. Methods: A systematic search of PubMed, Cochrane Library, Embase, and Web of Science was conducted for studies published between May 2017 and April 2024, adhering to PRISMA guidelines. Eligible studies were evaluated for quality utilizing the CLAIM and QUADAS-2 tools. Models trained on LiTS 2017 and 3DIRCADb datasets were included, with Dice Similarity Coefficient (DSC) employed as the primary performance metric. A Wilcoxon Signed-Rank Test was utilized to analyze performance distributions. Results: From an initial 224 studies, 41 met inclusion criteria. Among these, models trained on LiTS 2017 achieved a mean Dice Similarity Coefficient (DSC) of 0.846 (SD = 0.078), while those on 3DIRCADb attained 0.827 (SD = 0.071). Specialized architectures such as Spatial Attention Deep Supervised Network (SADSNet) and Multi-Attention Perception-Fusion U-Net (MAPFUNet) demonstrated superior performance compared to conventional models. Specialized architectures, including Convolutional Encoder-Decoder Residual Neural Network (CEDRNN) and Graph Convolutional Network (GCN), exhibited promising results but were limited to the LiTS 2017 dataset. Variability in performance was influenced by dataset characteristics, algorithm architecture, and evaluation protocols. Conclusions: Deep learning algorithms demonstrate high efficacy in liver tumor segmentation, achieving metrics comparable to manual segmentation. These findings substantiate the potential of artificial intelligence-powered imaging in enhancing cancer diagnosis and personalized treatment. However, challenges such as dataset variability and generalizability necessitate further investigation to fully integrate these tools into clinical practice.