Sample Size Estimation to Achieve Sufficient Diagnostic Performance of Deep Learning for Gastric Cancer Peritoneal Metastasis

Abstract

Abstract Peritoneal metastasis, the most common form of distant metastasis in gastric cancer, involves the dissemination of cancer cells throughout the peritoneal cavity. The Peritoneal Cancer Index (PCI) is widely used in Western countries to evaluate peritoneal dissemination and plays a crucial role in prognostic prediction. However, PCI scoring depends heavily on surgeons’ visual assessment of lesion morphology and size, introducing subjectivity and inter-rater variability. To address this, constructing a low-cost and objective evaluation model using deep learning is a promising approach. Nevertheless, deep learning typically requires large-scale datasets, and since PCI involves 13 intra-abdominal regions, collecting sufficient data can be challenging. In this study, we aim to estimate the sample size required to achieve adequate generalization performance by constructing learning curves based on a limited dataset. This approach may contribute to establishing an effective communication tool bridging medical and engineering disciplines.