[Paper Review] MAPPING: Model Average with Post-processing for Stroke Lesion Segmentation

Jiayu Huo, Liyun Chen+5 more|arXiv (Cornell University)|Nov 11, 2022

Acute Ischemic Stroke Management4 citations

TL;DR

This paper presents MAPPING, a model averaging with post-processing approach for stroke lesion segmentation using the ATLAS v2.0 dataset. By training four diverse nnU-Net models with different configurations and ensembling their predictions, followed by a tailored post-processing step, the method achieved the highest performance in the 2022 MICCAI ATLAS Challenge, with a Dice score of 0.6667 and a final rank of 1 on the hidden test set.

ABSTRACT

Accurate stroke lesion segmentation plays a pivotal role in stroke rehabilitation research, to provide lesion shape and size information which can be used for quantification of the extent of the stroke and to assess treatment efficacy. Recently, automatic segmentation algorithms using deep learning techniques have been developed and achieved promising results. In this report, we present our stroke lesion segmentation model based on nnU-Net framework, and apply it to the Anatomical Tracings of Lesions After Stroke (ATLAS v2.0) dataset. Furthermore, we describe an effective post-processing strategy that can improve some segmentation metrics. Our method took the first place in the 2022 MICCAI ATLAS Challenge with an average Dice score of 0.6667, Lesion-wise F1 score of 0.5643, Simple Lesion Count score of 4.5367, and Volume Difference score of 8804.9102. Our code and trained model weights are publicly available at https://github.com/King-HAW/ATLAS-R2-Docker-Submission.

Motivation & Objective

• To develop a robust, generalizable stroke lesion segmentation method capable of handling the high variability in multi-center MRI datasets.
• To address overfitting and distribution shift in deep learning models trained on heterogeneous stroke MRI data.
• To improve segmentation performance on small and complex lesions through model ensemble and post-processing.
• To achieve state-of-the-art results on the ATLAS v2.0 benchmark, particularly on the hidden test set.
• To provide a publicly available, reproducible solution for stroke lesion segmentation in clinical and research settings.

Proposed method

• Trained four distinct nnU-Net models using different data augmentation, loss functions, and training schemes to maximize prediction diversity.
• Applied model averaging by ensembling the predictions from the four trained models to improve robustness and reduce variance.
• Implemented a post-processing strategy that removes small isolated false positives and refines lesion boundaries using morphological operations and connected component analysis.
• Used size-stratified five-fold cross-validation on the training set to ensure balanced lesion size distribution across folds.
• Employed MNI-152 template normalization and registration to standardize input images across diverse acquisition protocols.
• Validated the method on both public and hidden test sets using the official MICCAI evaluation platform with strict metric weighting to discourage overfitting.

Experimental results

Research questions

• RQ1 Can model averaging across diverse training schemes improve generalization in stroke lesion segmentation on a multi-center MRI dataset?
• RQ2 To what extent can post-processing refine ensemble predictions to improve specific evaluation metrics without overfitting?
• RQ3 How does the method perform on small and unconnected lesions, which are known challenges in stroke segmentation?
• RQ4 Can a combination of diverse training schemes and post-processing outperform individual models on the hidden test set of the ATLAS v2.0 challenge?
• RQ5 Does the method maintain high performance across diverse scan qualities and lesion morphologies found in real-world clinical data?

Key findings

• The MAPPING method achieved the highest performance in the 2022 MICCAI ATLAS Challenge, securing 1st place on the hidden test set with a Dice score of 0.6667.
• The ensemble of four diverse models significantly improved segmentation accuracy, particularly in reducing false positives and smoothing lesion boundaries compared to individual models.
• The post-processing step effectively reduced small spurious regions and improved the Lesion-wise F1 score to 0.5643 and the Simple Lesion Count score to 4.5367.
• Despite strong overall performance, the method still struggled with very small lesions and unconnected components, as shown in qualitative failure cases.
• The method demonstrated robustness across diverse scan protocols and lesion morphologies, indicating strong generalization capability on unseen data.
• The final leaderboard ranking, with a weighted score of 3.317, confirmed the method's superiority over other competing approaches, including those with more complex architectures.