Ultra-Low-Power Edge Deployment of Compact Calibration-Aware Depthwise CNNs for Real-Time Point-of-Care Renal CT Triage on Embedded TPU and FPGA Architectures

• Ada John

  ladoke Akintola university of technology

  Author

Emergency department overcrowding and radiation exposure from non-contrast computed tomography (CT) for suspected renal colic remain persistent clinical challenges, particularly in resource-constrained settings where point-of-care triage solutions are urgently needed. Existing deep learning models for kidney CT classification prioritize accuracy over calibration and computational efficiency, rendering them unsuitable for deployment on ultra-low-power edge accelerators. This study addresses the critical gap between high-performance kidney CT classification and practical edge deployment by proposing a calibration-aware, depthwise-separable convolutional neural network (CNN) optimized for embedded Tensor Processing Unit (TPU) and Field Programmable Gate Array (FPGA) architectures. We leverage the NephroNet architecture—a compact 1.46M-parameter depthwise-separable CNN with squeeze-and-excitation channel attention and SpatialGate—originally validated on a patient-disjoint, multi-center dataset of 12,446 kidney CT images . Our methodology integrates post-hoc temperature scaling for probability calibration, INT8 quantization-aware training, and hardware-specific optimization for Google Edge TPU and low-power FPGA platforms. Under simulated edge deployment conditions, the quantized model achieves 89.4% accuracy (95% CI: 87.8–91.0%) with an Expected Calibration Error (ECE) of 0.021 after temperature scaling (T*=1.42), while consuming an estimated 7.8–12.4 mJ per inference on Edge TPU and maintaining sub-15ms latency on FPGA configurations . These results demonstrate that calibration-aware compact CNNs can enable real-time, radiation-sparing renal CT triage at the point of care, with significant implications for emergency medicine workflows in both high-volume and resource-limited clinical environments.

• pdf

Copyright (c) 2026 Ada John (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.