Scholarlink

<p>Convolutional Neural Network (CNN)-based segmentation models process every input at full depth, leading to high computational costs and limiting real-time use on edge or IoT devices. This work proposes a region-specialized multi-exit network with a dynamic exit policy for near real-time medical image segmentation. The model adapts inference based on the input complexity and anatomical subregion difficulty. A novel exit policy jointly optimizes segmentation precision (via 95th-percentile-Hausdorff Distance) and computational cost (FLOPs), using local uncertainty cues and efficiency scores with a global Pareto frontier of latency, energy, and accuracy trade-offs. A region-specialized supervision mechanism further enhances the learning efficiency, where early exits handle simpler anatomical structures, while deeper exits refine complex subregions. The method is extensively evaluated on BraTS2020 and KiTS19, achieving high segmentation accuracy while enabling dynamic routing that serves as an asynchronous load-balancer for HPC environments and delivers true real-time throughput on constrained edge hardware.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fast and efficient inference strategy for medical image segmentation using input-aware dynamically configured neural network architecture

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Binit Kumar Pandit,
  • Ayan Banerjee

摘要

Convolutional Neural Network (CNN)-based segmentation models process every input at full depth, leading to high computational costs and limiting real-time use on edge or IoT devices. This work proposes a region-specialized multi-exit network with a dynamic exit policy for near real-time medical image segmentation. The model adapts inference based on the input complexity and anatomical subregion difficulty. A novel exit policy jointly optimizes segmentation precision (via 95th-percentile-Hausdorff Distance) and computational cost (FLOPs), using local uncertainty cues and efficiency scores with a global Pareto frontier of latency, energy, and accuracy trade-offs. A region-specialized supervision mechanism further enhances the learning efficiency, where early exits handle simpler anatomical structures, while deeper exits refine complex subregions. The method is extensively evaluated on BraTS2020 and KiTS19, achieving high segmentation accuracy while enabling dynamic routing that serves as an asynchronous load-balancer for HPC environments and delivers true real-time throughput on constrained edge hardware.