State-of-the-art TinyML approaches for colorectal cancer detection: Current advances, challenges, and future directions.

Abstract

Colorectal cancer (CRC) remains a leading cause of cancer-related mortality worldwide, with diagnostic disparities, particularly pronounced in resource-constrained and decentralized healthcare settings. Recent advances in TinyML machine learning models optimized for ultra-low-power, memory-constrained embedded devices have created new opportunities for scalable on-device CRC screening and diagnostics. This review presents a systematic and CRC-centric analysis of TinyML technologies across the diagnostic continuum, including capsule endoscopy, histopathology, breath analysis, and biosignal-based screening. Unlike existing surveys that address TinyML from a general healthcare perspective, this study focuses specifically on the technical, clinical, and deployment challenges unique to CRC diagnostics. We propose a structured taxonomy encompassing model compression techniques, hardware-software co-design strategies, and clinical deployment paradigms, and critically analyze the accuracy-latency-energy trade-offs across representative platforms. This review further synthesizes recent (2024-2025) advances in TinyML compilers, hardware accelerators, and edge-cloud integration, highlighting their implications for real-world clinical translation. By consolidating current evidence, identifying benchmarking and regulatory gaps, and outlining a forward-looking research roadmap, this survey clarifies the role of TinyML as a viable enabler of real-time, privacy-preserving, resource-efficient CRC diagnostics. These findings provide actionable insights for researchers, clinicians, and system designers seeking to deploy TinyML solutions in equitable and clinically meaningful cancer care.