Intracranial aneurysm segmentation with nnU-net: utilizing loss functions and automated vessel extraction

TL;DR

This paper introduces a new method for accurately segmenting intracranial aneurysms in MRI scans using advanced neural networks and loss functions.

Contribution

The novel integration of hybrid loss functions and automated vessel extraction improves segmentation accuracy for intracranial aneurysms.

Findings

The model achieved a Dice coefficient of 0.72 on the RENJI dataset and 0.54 on the ADAM dataset.

Sensitivity scores reached 0.69 on RENJI and 0.53 on ADAM, showing improved detection performance.

Combining vessel information with hybrid loss functions enhanced the model's focus on critical aneurysm regions.

Abstract

Intracranial aneurysms pose significant challenges in diagnosis and treatment, emphasizing the need for accurate segmentation methods to assist clinicians in their management. In this paper, we present a novel approach for segmenting intracranial aneurysms using three-dimensional time-of-flight magnetic resonance angiography (TOF-MRA) images and the no-new-U-net framework. We aim to improve segmentation accuracy and efficiency through the integration of hybrid loss functions and additional vessel information.

The model was conducted on Aneurysm Detection And SegMentation (ADAM) and Renji Hospital (RENJI) datasets. The TOF-MRA ADAM dataset contains data from 113 cases, where 89 have at least one aneurysm with a median maximum diameter of 3.6 mm and range from 1.0 to 15.9 mm. The RENJI private TOF-MRA dataset comprises 213 cases including both ruptured and unruptured aneurysms with a median maximum diameter of 9.35 mm (range: 1.25-37.58 mm). We optimized the segmentation model by exploring hybrid loss functions that combine distribution-based and region-based losses to effectively delineate intricate aneurysm structures. Additionally, we incorporated vessel information as a region of interest using an automatic vessel segmentation algorithm to enhance the model's focus on critical regions. The model was trained on multi-modality data, including both vessel-enhanced and original images, to capture complementary information and improve segmentation accuracy.

Extensive simulations on both the ADAM dataset and a private RENJI dataset demonstrate the effectiveness of our approach. The best-performing loss function yielded significant improvements in the Dice coefficient (0.72 and 0.54) and Sensitivity (0.69 and 0.53) on the RENJI and ADAM datasets, respectively.

The proposed method offers a promising solution for accurately segmenting intracranial aneurysms, showcasing superior performance compared to existing approaches. By integrating hybrid loss functions and vessel information, we enhance the model's ability to delineate intricate aneurysm structures, contributing to improved diagnosis and treatment planning for patients with intracranial aneurysms.

Figures