نهمین کنگره نقشه برداری مغز

نمایش مقاله

Title:

Automated MS Lesion Segmentation Using Self-Configuring nnU-Net Framework: Evaluation on MSSEG2016 Dataset

Subject:

Clinical Applications of Brain Mappin

Presenter Author:
Name:	Zahra Arab

Author:
Name 1:	Zahra Arab - Presenter
Affiliation 1:	Department of Biomedical Engineering, SR.C., Islamic Azad University, Tehran, Iran

Name 2:	Mahsa Akhbari - Corresponding
Affiliation 2:	Department of Biomedical Engineering, SR.C., Islamic Azad University, Tehran, Iran

Name 3:
Affiliation 3:

Name 4:
Affiliation 4:

Name 5:
Affiliation 5:

Name 6:
Affiliation 6:

Name 7:
Affiliation 7:

Name 8:
Affiliation 8:

Name 9:
Affiliation 9:

Name 10:
Affiliation 10:

Abstract

Original Research

introduction :

Multiple Sclerosis (MS) is a prevalent neurodegenerative disease affecting 2.8 million individuals globally [1]. Precise White Matter Lesion (WML) segmentation on MRI is critical for diagnosis (McDonald criteria [2]) and monitoring treatment response. However, manual segmentation is labor-intensive (20-60 minutes/case) and subject to inter-rater variability [3]. While Deep Learning models achieve high accuracy, they require extensive manual tuning and expertise, limiting clinical adoption [4, 5]. This study evaluates nnU-Net [6], a self-configuring framework automating the entire segmentation pipeline, enabling rapid deployment in non-specialized centers.

Methods:

We utilized the nnU-Net framework, which automates architecture configuration based on dataset properties. The pipeline includes: (1) Fixed parameters for generalization (standard U-Net, SGD optimizer with Nesterov momentum μ=0.99); (2) Rule-based parameters (dynamic adaptation of patch size and resampling based on image dimensions and voxel spacing); and (3) Empirical parameters derived from cross-validation (model ensemble selection, post-processing). The study utilized the MSSEG2016 dataset [7] comprising 15 training and 38 testing cases acquired from multiple centers (Siemens, Philips, General Electric; 1.5T and 3T). During quality control, two test cases were excluded: one due to severe image contrast degradation rendering tissue boundaries indistinguishable, and another due to gross anatomical abnormalities incompatible with the training distribution. To handle class imbalance (small lesions versus extensive background), we employed a combined loss function: L_total=L_Dice+L_CE where L_Dice measures region overlap and L_CE (cross-entropy loss) penalizes pixel-wise classification errors. The Dice loss is defined as: L_Dice=1-(2∑_(i=1)^N▒p_i ⋅g_i)/(∑_(i=1)^N▒p_i +∑_(i=1)^N▒g_i ) where p_i represents the predicted probability for voxel i (0-1 range), g_i denotes the ground truth binary label, and N is the total voxel count. The cross-entropy component ensures gradient stability. Performance was assessed using Dice Similarity Coefficient (DSC), Sensitivity, and Positive Predictive Value (PPV).

Results:

Quantitative evaluation on 36 test cases is presented in Table 1. Mean±SD and median values were: Dice 0.741±0.134 (median 0.777), Sensitivity 0.757±0.189 (median 0.792), and PPV 0.749±0.173 (median 0.781). The lesion-free case was correctly identified with zero false positives (perfect specificity, Dice=1.0). Three representative cases demonstrating varying segmentation quality are shown in Figure 1 (left to right columns: automatic segmentations, original FLAIR images): Case S-49 achieved Dice=0.892, Case S-17 achieved Dice=0.846, and Case S-27 achieved Dice=0.778. Model training required 20 hours. Inference time averaged 3 minutes per case, enabling real-time clinical integration. Recent manually-optimized methods report higher accuracy but require extensive tuning and prolonged training [4, 5]. Table 1: Performance Metrics on MSSEG2016 Test Set (N=36) Metric Mean ± SD Median Dice Coefficient 0.741 ± 0.134 0.777 Sensitivity 0.757 ± 0.189 0.792 Positive Predictive Value 0.749 ± 0.173 0.781 Processing Time (min/case) 3.1 ± 0.4 3.0

Conclusion:

This study demonstrates that nnU-Net provides a robust, "out-of-the-box" solution for MS lesion segmentation without manual architecture tuning. The primary novelty lies in democratizing state-of-the-art segmentation: institutions lacking deep learning expertise can deploy clinically viable models immediately. Key advantages include: (1) zero manual configuration, (2) 87-95% time reduction versus manual segmentation, (3) consistent performance eliminating inter-rater variability, and (4) seamless cross-institutional transferability. While the model excels on medium-to-large lesions, future work will focus on improving sensitivity for micro-lesions through integration of attention mechanisms and multi-scale feature fusion. The method's viability for clinical workflow integration and automated diagnostic assistants is confirmed.

Keywords:

Multiple Sclerosis, Lesion Segmentation, nnU-Net, MSSEG2016 Dataset, MRI Images