Bilkent University
Department of Computer Engineering
CS 590/690 SEMINAR
GPU-Accelerated Topology-Preserving Loss Functions for Biomedical Image Segmentation
Ahmet Caner Akar
Master Student
(Supervisor:Asst.Prof.Doruk Öner)
Computer Engineering Department
Bilkent University
Abstract: Preserving the topological integrity of curvilinear anatomical structures, such as blood vessels and neurons, is critical for biomedical image segmentation. While topology-aware loss functions such as CAPE, MALIS, and Persistent Homology effectively address this challenge, their high computational cost severely limits their practical adoption. In this work, we analyze the computational bottlenecks of these loss functions and develop custom CUDA kernels to accelerate their execution on GPUs. Specifically, we implement a parallel delta-stepping Dijkstra algorithm for shortest-path computation, which forms the core of the CAPE loss, integrate it with PyTorch's autograd framework, and present comprehensive performance benchmarks against CPU baselines. Our preliminary results demonstrate that shortest path computation accounts for over 60% of total CPU execution time, and GPU parallelization has significant potential to eliminate this bottleneck, enabling topology-preserving models to be trained on large-scale 3D biomedical datasets within practical timeframes.
DATE: April 20, Monday @ 16:30 Place: EA 502