Fuzzy-logic enhanced internet of things model for uncertainty-adaptive medical image segmentation

Abstract

Deep learning (DL) models for brain tumour segmentation (BTS) typically produce a fixed decision boundary and do not allow the uncertainty they estimate to impact the segmentation decision (SD). This work presents a FUZIONet-Med: A Fuzzy Uncertainty-Adaptive IoT Edge Intelligence Model for Brain Tumor Segmentation in which decomposed voxel-level uncertainty governs the SD within a single network. Monte Carlo dropout (MCD) generates epistemic and aleatoric uncertainty maps, which drive a Mamdani fuzzy inference system through an asymmetric nine-rule base to produce a voxel-level adaptive threshold field. This field modulates the encoder skip connections through a fuzzy-guided attention mechanism (FGAM) and, at inference, replaces the conventional fixed threshold as the voxel-level decision boundary, so that the segmentation boundary varies with local uncertainty. The model is deployed within a three-tier IoT architecture in which edge nodes (EN) perform acquisition, pre-processing, and inference, and a practical Byzantine fault-tolerant (PBFT)-secured protocol governs federated aggregation. On the BraTS 2013 benchmark under repeated stratified cross-validation, the model attained Dice scores of 0.908, 0.851, and 0.769 for the whole tumour (WT), tumour core (TC), and active tumour (AT) regions, with a 95th-percentile Hausdorff distance of 10.14 mm and an uncertainty calibration error (UCE) of 0.121, exceeding 6 baselines on every metric with Holm-adjusted statistical significance. In this study, we performed an external evaluation on 1,221 independent BraTS 2021 subjects using a model configuration fixed exclusively to the BraTS 2013 development protocol, without retraining, fine-tuning, or external-cohort-driven parameter optimization. Full-resolution inference was completed in 2930 ms at 28.9 W on an NVIDIA Jetson AGX Xavier EN, within a 30 W envelope. The results indicate that coupling decomposed uncertainty to the inference-time decision boundary through in-network fuzzy control improves segmentation accuracy and calibration while remaining deployable on clinical edge hardware.