损伤Synth:一种简单的参数损伤合成方法,用于在低数据场景中改善脊髓损伤细分
Ricky Walsh1, Prabhjot Kaur2, Davood Karimi2
1Univ Rennes, Inria, CNRS, Inserm, IRISA UMR 6074, Empenn, Rennes, France.
Imaging neuroscience (Cambridge, Mass.)
|December 1, 2025
概括
使用LesionSCynth在脊髓MRI中合成现实的多发性硬化症 (MS) 病变,显著提高了深度学习模型的性能. 这种方法减少了对广泛的手册注释的需求,提高了在低数据场景中的病变检测和细分.
科学领域:
- 医疗成像医学成像
- 人工智能的人工智能
- 神经学 神经学
背景情况:
- 在脊髓MRI中检测多发性硬化症 (MS) 病变至关重要,但具有挑战性.
- 深度学习模型需要大量的注释数据集,这些数据集的生产成本高,耗时长.
研究的目的:
- 引入LesionSCynth,用于在脊髓MRI中合成高强度MS病变的框架.
- 为了减少训练深度学习模型的注释负担.
主要方法:
- LesionSCynth分析T2加权MRI中的真实病变强度分布,以产生合成病变.
- 该框架增加了这些合成损伤的小注释数据集.
- 细分模型使用真实数据和合成数据的组合进行训练.
主要成果:
- 使用LesionSCynth训练的模型表现出更好的性能 (0.52 FROC) 与仅在真实数据 (0.46 FROC) 上训练的模型相比.
- 合成病变的表现与在八倍多的真实数据 (0.55 FROC) 上训练的模型相匹配.
- 在低数据条件下,LesionSCynth的性能优于LesionMix和CarveMix等现有方法.
结论:
- LesionSCynth是一种实用和有效的工具,用于增强在脊髓MRI中的MS病变检测和细分.
- 该框架大大降低了深度学习模型培训的注释要求.
相关概念视频
01:00Cellular Injury I: Introduction
Cellular injury occurs when a cell cannot maintain homeostasis or adapt to stressors such as hypoxia, toxins, or trauma. Depending on severity and duration, injury may be reversible, allowing recovery, or irreversible, leading to cell death.General Mechanisms of Cell InjuryAlthough causes vary, most cellular injuries arise from a few key mechanisms that disrupt essential functions and often amplify one another. Cell survival depends on the extent and balance of these disturbances.ATP depletion...
01:21Cellular Injury II: Classification
Cellular injury is any process that disrupts a cell’s ability to maintain homeostasis, leading to structural or functional changes. It is broadly classified based on etiology (cause) and mechanism of damage.Classification by EtiologyCellular injury may result from several causes. Hypoxic injury happens due to reduced oxygen delivery, most commonly from inadequate blood supply, such as arterial obstruction; for example, coronary artery thrombosis can cause myocardial infarction. Chemical injury...