The Accuracy Paradox: A Systematic Review of the Computational Reporting Gap in Deep Learning for Breast Cancer Imaging

Abstract

Deep learning has transformed breast cancer imaging by substantially improving automated lesion detection, classification, and diagnostic decision support. The PRISMA-compliant systematic review investigates the computational performance and reporting transparency of DL models in breast cancer imaging (2021–2025). From an initial 3,748 Scopus records, 65 high-quality studies were synthesized, identifying Mammography, Ultrasound, MRI, Thermography, and Histopathology as the primary modalities. Architecturally, the landscape reveals a transition from standard Convolutional Neural Networks (CNNs) toward sophisticated frameworks, including Vision Transformers (ViT), YOLOv5-CBAM, and Generative Adversarial Networks (GANs). While diagnostic accuracy is frequently reported at near-perfect levels (up to 99.9%), this review uncovers a pervasive Accuracy Paradox intertwined with a critical computational reporting gap. Specifically, over 76.92% of the literature fails to document essential technical metrics such as Floating Point Operations (FLOPs), inference latency, or memory consumption. Although pioneering studies utilize FPGA hardware acceleration and quantization-aware training, standardized efficiency reporting remains urgent. This review concludes that balancing predictive power with transparent computational overhead is essential for the successful translation of these "black-box" models into sustainable, real-time clinical engineering environments.