Resource-efficient Medical Image Segmentation Based on Self-supervised Learning and Dynamic Multimodal Sensor Fusion

Rapid advancements in high-definition CMOS and magnetic resonance transducers have led to the accumulation of complex medical imaging data that requires robust, real-time computational interpretation.However, current high-performance segmentation models require excessive computational power, making them incompatible with low-power point-of-care sensing hardware.Therefore, we improved the Self-Supervised Dynamic Gated Fusion Network (SS-DGFNet) model for resource-efficient medical image segmentation.The network utilizes automated signal calibration (self-supervised learning) and an adaptive fusion module to maintain high accuracy even with missing sensor data or limited labeled information.For the Multimodal Brain Tumor Image Segmentation Benchmark 2025 dataset, SS-DGFNet shows high spatial accuracy (a Dice score of 0.888) while maintaining 97.6% performance retention when a sensor channel is lost.Despite these gains, issues remain, including the need for validation across a broader range of clinical sensor materials and the optimization of the model for heterogeneous edge-computing hardware.The improved model demonstrates significant robustness when a sensor modality is missing.By reducing computational overhead and accelerating calibration cycles for emerging biosensors, the model leads to the transition of complex diagnostics to edge-computing sensor platforms and supports the transition of complex diagnostics to mobile sensor platforms.