Rectal Cancer Segmentation: A Methodical Approach for Generalizable Deep Learning in a Multi-Center Setting

Abstract

Noninvasive Artificial Intelligence (AI) techniques have shown great potential in assisting clinicians through the analysis of medical images. However, significant challenges remain in integrating these systems into clinical practice due to the variability of medical data across multi-center databases and the lack of clear implementation guidelines. These issues hinder the ability to achieve robust, reproducible, and statistically significant results. This study thoroughly analyzes several decision-making steps involved in managing a multi-center database and developing AI-based segmentation models, using rectal cancer as a case study. A dataset of 1212 Magnetic Resonance Images (MRIs) from 14 centers was used. The study examined the impact of different image normalization techniques, network hyperparameters, and training set compositions (in terms of size and construction strategies). The findings emphasize the critical role of image normalization in reducing variability and improving performance. Additionally, the study underscores the importance of carefully selecting network structures and loss functions based on the desired outcomes. The potential of clustering approaches to identify representative training subsets, even with limited data sizes, was also evaluated. While no definitive preprocessing pipeline was identified, several networks developed during the study produced promising results on the external validation set. The insights and methodologies presented may help raise awareness and promote more informed decisions when implementing AI systems in medical imaging.