A Holistic Weakly Supervised Approach for Liver Tumor Segmentation with Clinical Knowledge-Informed Label Smoothing

Liver is also a common destination for metastatic cancer cells originating from various abdominal organs, including the colon, rectum, pancreas, as well as distant organs such as the breast and lung. Consequently, a thorough examination of the liver and its lesions is critical to comprehensive tumor staging and management strategies. Standard tumor assessment protocols, such as the Response Evaluation Criteria in Solid Tumor (RECIST), require precise measurement of the diameter of the largest target lesion (Eisenhauer et al., 2009). Thus, accurate localization and precise segmentation of liver tumors within CT scans are essential for effective diagnosis, treatment planning, and monitoring of therapeutic response in patients with liver cancer (Shiina et al., 2018; Terranova & Venkatakrishnan, 2024; Virdis et al., 2019). Manual delineation of target lesions in CT scans is fraught with challenges, being both time-consuming and prone to poor reproducibility and operator-dependent variability (Gul et al., 2022). Automated liver tumor segmentation can provide clinicians with rapid and consistent tumor delineation, thereby improving patient outcomes and reducing healthcare costs. Recently, deep learning algorithms have shown promise for producing automated liver and tumor segmentation (Gul et al., 2022). While many algorithms achieved exceptional performance in liver segmentation, with dice scores ranging from 0.90 to 0.96, enhancing liver tumor segmentation remains a challenge, currently standing at dice scores from 0.41 to 0.67 according to a recent Liver Tumor Segmentation Benchmark (Bilic et al., 2023). Liver tumor segmentation is an inherently challenging task because tumors vary significantly in size, shape, and location across different patients, which leads to a broad spectrum of tumor characteristics and hinders model generalization (Sabir et al., 2022).