Year: 2025
Type: dissertation
Abstract: Semantic segmentation of medical images is a critical task in modern healthcare, enabling precise identification and localization of anatomical structures and pathological regions by classifying image regions at the pixel level, thereby supporting rapid and accurate diagnosis. Although deep learning models have achieved remarkable performance in this domain, their deployment in real-world clinical environments is often constrained by the high computational demands of traditional software implementations, resulting in inefficiency and latency. Field-Programmable Gate Arrays (FPGAs) provide a promising alternative by offering programmable hardware-level acceleration tailored for deep learning inference tasks, with advantages such as parallel processing, low latency, and energy-efficient computation, making them particularly suitable for time-sensitive applications including tumor detection, organ delineation, and lesion classification. In this work, a lightweight neural network architecture, Mobile-CMUNeXt, was designed and optimized for FPGA deployment through architectural pruning, quantization-aware training, and hardware-friendly modifications, alongside the development of modular accelerator cores for depthwise, pointwise, and 3D convolutions implemented on the Avnet Ultra96-V2 platform. Experimental results demonstrate that the proposed accelerator achieves real-time inference performance with clinically acceptable segmentation accuracy, while balancing resource utilization and maintaining energy efficiency and low latency, and further show competitive throughput compared to CPU and GPU baselines, validating the effectiveness of the hardware–software co-design methodology. Overall, this thesis establishes that FPGA-based accelerators constitute a viable solution for deploying deep learning models in edge medical devices by combining efficient neural network design with custom hardware optimizations, thereby enabling real-time, privacy-preserving, and energy-efficient semantic segmentation of medical images.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7154922304