SKIN-ORBIT: A bio-mimetic oscillatory resonance-based inference topology for universal skin lesion segmentation

Abstract

Accurate segmentation of skin lesions is a foundational task in dermatology, essential for early diagnosis, treatment planning, and long-term monitoring of skin conditions. However, current state-of-the-art models—primarily based on convolutional neural networks and transformers—depend on rigid feature hierarchies and large-scale annotated datasets. These architectures struggle to generalize across diverse lesion morphologies, rare dermatological presentations, and variations in skin tone or imaging modality. In this work, we introduce SKIN-ORBIT (Spectral-Kinetic Inference Network with Oscillatory Resonance-Based Inference Topology), a biologically inspired, signal-centric segmentation framework that redefines how spatial patterns are modeled. Rather than relying on fixed receptive fields or attention over patches, SKIN-ORBIT treats each pixel as a localized oscillatory emitter, producing dynamic waveform signals over time. These signals propagate through a deformable resonance field, where zones of persistent phase coherence emerge as lesion regions—identified via stable, constructive interference patterns. The framework is anchored by two core components: the Kinetic Deformation Unit (KDU), which modulates spatial topology through dynamic elastic distortions driven by pixel-wise kinetic energy profiles, enabling continuous and topology-aware field reshaping; and the Oscillatory Field Propagation (OFP) module, which replaces convolution and attention with harmonic signal propagation, simulating wavefront interactions across the spatial domain without predefined kernels. To facilitate unsupervised learning, we propose a novel Cross-Modal Resonance Coherence Loss (CMRCL), which enforces phase alignment across visible, ultraviolet (UV), and infrared (IR) imaging modalities. CMRCL computes spectral-phase divergence in the latent oscillatory domain and minimizes incoherence between modality-specific representations, encouraging shared structural resonance. SKIN-ORBIT was trained and validated on a dataset of 15,318 cross-spectrum dermatological images. It achieved 98.7% segmentation accuracy, a Dice coefficient of 0.92, Jaccard index of 0.86, sensitivity of 0.91, specificity of 0.97, and a Hausdorff Distance of 5.6 pixels—substantially outperforming traditional CNN and transformer-based models. These results demonstrate that SKIN-ORBIT provides a disruptive alternative to conventional architectures, introducing a physics-driven, annotation-efficient paradigm for robust and modality-agnostic skin lesion segmentation.