Non‐Invasive Diabetes Detection Through Human Breath Using Hybrid Octave‐ <scp>CenterNet</scp> Neural Network With <scp>DenseNet</scp> ‐77 Model

ABSTRACT Diabetes Mellitus (DM), including Type 1 and Type 2, is a metabolic disorder caused by defects in insulin secretion or action. Non‐invasive detection is more critical because invasive methods often lack data and have reduced accuracy, leading to poorer machine learning performance. This research proposes a new Octave‐CenterNet with DenseNet‐77 framework for efficient detection and classification of diabetes from Volatile Organic Compounds (VOCs). The method combines a rapid discrete curvelet transform with wrapping to capture prominent features quickly, uses octave convolution to preserve high and low‐frequency patterns and enrich representations, employs CenterNet to detect acetone as a major biomarker, and leverages DenseNet‐77 for gradient‐efficient classification. Willow sled catkin optimization adaptively fine‐tunes hyperparameters to further enhance performance. The model effectively distinguishes healthy individuals from diabetic patients and differentiates between Type 1 and Type 2 diabetes. Experimental results demonstrate excellent performance with 98.7% accuracy, 98% precision, 99.7% recall, and 99.34% F1 score, validating its robustness. Overall, this end‐to‐end, noise‐resistant, and computationally efficient framework offers a technically advanced and practical solution for non‐invasive diabetic detection.