Optimizing nnU-Net with OpenVINO for Fast CPU Inference in Abdominal Organ Segmentation

Yannick Kirchhoff, Ashis Ravindran, Maximilian Rouven Rokuss, Benjamin Hamm, Constantin Ulrich, Klaus Maier-Hein, Fabian Isensee

Keywords: FLARE Challenge, Organ Segmentation, nnU-Net, OpenVINO

Abstract: Accurate segmentation of abdominal organs in pathological computed tomography (CT) scans is crucial for diagnosis and treatment planning. However, this task is challenging due to the diversity of organ appearances and sizes, as well as the computational limitations in clinical settings. Task 2 of the FLARE 2024 challenge was launched to encourage the development of algorithms capable of efficient abdominal organ segmentation under strict resource constraints, specifically focusing on CPU-based inference without access to GPUs. In this paper, we describe our contribution to this challenge by utilizing nnU-Net with optimizations for efficient CPU-based inference using OpenVINO. We resampled the CT scans to an isotropic low resolution to balance segmentation accuracy and computational efficiency. Our method achieved an average Dice Similarity Coefficient (DSC) of 76.8\% and an average Normalized Surface Dice (NSD) of 80.5\% on the public validation set, with an average running time of 26 seconds per case. These results demonstrate that our approach effectively addresses the challenges of efficient organ segmentation under resource constraints, and underscore the potential for deploying such methods in real-world clinical environments where computational resources are limited.

Submission Number: 21