UC Irvine

Photographic imaging is a vital tool for cancer detection and screening, enabling non-invasive visualization of abnormal tissues and early-stage malignancies. It is particularly effective for detecting and screening cancers that are visually accessible, such as skin, oral, cervical, and certain gastrointestinal cancers. Techniques such as dermoscopy, intraoral imaging, and endoscopy leverage high-resolution photographic methods to capture detailed views of body tissues. A key advantage of photographic imaging lies in its simplicity and the feasibility of constructing compact, low-cost systems, making it especially well-suited for large-scale cancer screening programs. However, this streamlined approach can lead to variability in image quality due to factors such as lighting conditions, focus, and operator handling. These inconsistencies pose challenges for AI-based diagnostic systems, where AI model performance can be highly sensitive to input variations. Consequently, ensuring the robustness of AI algorithms across diverse and degraded image conditions is essential for real-world deployment. In this study, we comprehensively evaluate the performance and robustness of state-of-the-art convolutional neural networks (CNNs), transformers, and general-purpose foundation models under common system-induced colorimetric/tonal and capture-induced degradations, and when degradation type combines. These degradations were carefully designed to simulate realistic image quality issues frequently encountered in clinical and screening settings of photographic imaging. Our findings provide a solid foundation and actionable insights for developing reliable deep learning classification systems in practical medical environments. We identified how training strategies, neural network architectures, and image degradation types individually and jointly influence AI-based diagnostic performance in photographic cancer detection. Furthermore, by accurately assessing how variations in photographic imaging devices and capture-induced image-quality changes affect AI-integrated diagnostic performance, this work highlights critical considerations for building future AI-integrated imaging systems that are resilient to real-world operational challenges.