Comparative evaluation of training strategies using partially labelled datasets for segmentation of white matter hyperintensities and stroke lesions in FLAIR MRI

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

This study evaluated six deep learning training strategies for segmenting white matter hyperintensities (WMH) and ischaemic stroke lesions (ISL) in FLAIR MRI scans using partially labelled datasets. WMH and ISL are visually confounding features associated with cerebral small vessel disease. Researchers combined privately held and publicly available datasets, totaling 2052 MRI volumes, with 1341 containing WMH annotations and 1152 containing ISL annotations. The investigation found that several methods effectively leveraged partially labelled data to enhance model performance, with pseudolabels yielding the best overall results. The models were trained and tested across 12 diverse datasets, demonstrating improved performance, particularly on datasets not fully represented in the fully labelled subset. Qualitative analysis revealed that pseudolabel model outputs were often as plausible as ground truth, even identifying missing deep WMH instances.

Key takeaway

For Machine Learning Engineers developing medical image segmentation models with incomplete annotations, you should prioritize pseudolabeling strategies. This approach significantly improves performance for complex, visually confounding lesions like WMH and ISL, especially when integrating diverse datasets. Consider implementing a pseudolabel-trained multiclass model to enhance accuracy and robustness, but be aware of its potentially higher training complexity compared to simpler loss-based methods. Your qualitative analysis of model outputs will be crucial to validate clinical utility.

Key insights

Pseudolabeling is the most effective strategy for deep learning segmentation of confounding medical lesions using partially labeled MRI data.

Principles

• Partially labelled data boosts model performance.
• Pseudolabels improve segmentation accuracy.
• WMH segmentation has high inter-rater variability.

Method

Six strategies were compared using DynUNet on 2052 FLAIR MRI volumes. Preprocessing included bias correction, resampling to 1x1x1 mm, brain extraction, and z-score standardization. Training used Dice + cross-entropy loss with SGD and extensive augmentation.

In practice

• Use pseudolabels for optimal partially supervised segmentation.
• Discard unclear ISL scans from training/test sets.
• Employ DSCθ for robust WMH segmentation evaluation.

Topics

• Medical Image Segmentation
• Partially Labelled Data
• White Matter Hyperintensities
• Ischaemic Stroke Lesions
• Deep Learning
• FLAIR MRI
• Pseudolabeling

Code references

• Jesse-Phitidis/partially_labelled_WMH_ISL_seg

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Attention-Based Segmentation of WMHs and Differentiation of Vascular vs. Demyelinating Lesions — Combining attention-based segmentation with feature-driven classification improves differentiation of WMH etiologies in MRI.
• GLI-AL: A Multi-Modal Glioma MRI Label Resource with Unified Anatomy-Lesion Labels — GLI-AL unifies anatomy-lesion labels for glioma MRI, improving segmentation by including white matter hyperintensities.
• BrainG3N: A Dual-Purpose Tokenizer for Controllable 3D Brain MRI Generation — A novel dual-purpose tokenizer enables a unified 3D brain-MRI embedding space for clinical tasks and controllable generation.
• Mutual Distillation of Dual-Foundation Models for Semi-Supervised PET/CT Segmentation — Mutual distillation of dual-foundation models improves semi-supervised PET/CT segmentation with minimal labeled data.
• Temporally Consistent Label Interpolation for Robust Surgical Multi-Task Learning under Challenging Conditions — FAROS interpolates sparse surgical annotations into dense, temporally consistent labels using optical flow, balancing multi-task learning.
• Improving Pre-trained Adult Glioma Segmentation Models Using only Post-processing Techniques — Adaptive, radiomic-based post-processing significantly enhances pre-trained glioma segmentation model accuracy without additional GPU resources.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.