具跨領域少樣本適應能力之磁振造影運動偽影檢測器

Abstract

磁振造影（Magnetic Resonance Imaging, MRI）廣泛應用於臨床診斷，然而病患運動產生之運動偽影會降低影像品質，並影響後續人工智慧與定量分析之可靠性。此外，不同器官、掃描協定及影像特性造成之資料域差異，可能降低模型於新資料上的泛化能力，增加資料標註與模型建立成本。本研究旨在建立具跨域泛化與少樣本適應能力之MRI運動偽影檢測器，並探討模型最佳化、跨域適應及模型預測與影像紋理特徵之關聯。 本研究採用三個高度異質之MRI資料集，包括公開膝關節資料集KMAR-50K、腦部資料集MR-ART，以及雙和醫院蒐集之肝臟資料集Liver-MRI。首先透過消融實驗評估骨幹網路、切片取樣策略及病人層級聚合方法，以建立最佳化訓練流程；其後分別訓練來源模型並進行跨域驗證，再以少量標註資料進行少樣本適應，以評估模型之跨域泛化與性能恢復。最後，以規則式方法及GLCM結合傳統分類器作為基準，並分析模型排序切片之紋理變化。 結果顯示，Swin-Tiny具有最佳整體分類性能，不同來源模型呈現非對稱的跨域泛化能力。其中，MR-ART來源模型於外部目標資料域進行20-shot微調後，可恢復超過90%的完整監督訓練性能，展現較佳的跨域適應能力。傳統紋理特徵方法之性能低於深度學習模型，顯示模型可利用紋理以外的影像資訊判別運動偽影。GLCM分析亦發現，實驗組模型排序前三之切片具有較明顯的紋理變化。 本研究建立整合模型最佳化、跨域泛化、少樣本適應及模型決策驗證之MRI運動偽影檢測架構。研究結果顯示，少量目標域資料即可恢復大部分模型性能，可望降低新醫療場域建立運動偽影檢測模型之標註成本，並作為MRI自動化品質控制與人工智慧影像分析系統跨場域部署之基礎。

Magnetic resonance imaging (MRI) is widely used in clinical diagnosis; however, motion artifacts caused by patient movement can degrade image quality and affect the reliability of subsequent artificial intelligence (AI)-based and quantitative analyses. Domain shifts arising from differences in anatomical regions, imaging protocols, and image characteristics may further reduce model generalizability to new datasets, increasing the costs of data annotation and model development. This study aimed to develop an MRI motion artifact detector with cross-domain generalization and few-shot adaptation capabilities, while investigating model optimization, cross-domain adaptation, and the relationship between model predictions and image texture characteristics. Three highly heterogeneous MRI datasets were used: the public knee dataset KMAR-50K, the public brain dataset MR-ART, and an in-house liver dataset Liver-MRI collected at Taipei Medical University Shuang Ho Hospital. Ablation experiments were first conducted to evaluate backbone architectures, slice-level sampling strategies, and patient-level aggregation methods to establish an optimized training pipeline. Source models were then trained on each dataset and evaluated through cross-domain validation, followed by few-shot adaptation using limited labeled target domain data. A rule-based method and GLCM-based conventional classifiers were also used as baselines, and GLCM analysis was performed to examine texture changes in model-ranked slices. Swin-Tiny achieved the best overall classification performance, while different source models exhibited asymmetric cross-domain generalization. The MR-ART source model achieved the best overall adaptation performance, with more than 90% of the fully supervised target domain performance recovered after 20-shot fine-tuning on the external target domains. Conventional texture-based methods generally performed worse than the deep learning model, suggesting that the evaluated GLCM descriptors alone were insufficient to reproduce the classification performance of the learned model. GLCM analysis further showed systematic texture differences between the Top-ranked and Remaining slices within motion artifact examinations. In conclusion, this study established an MRI motion artifact detection framework integrating model optimization, cross-domain generalization, few-shot adaptation, and model decision analysis. The results demonstrate that a small amount of target domain data can recover most of the performance of fully supervised training, potentially reducing annotation costs when deploying motion artifact detection models in new clinical environments. The proposed framework may provide a basis for cross-domain deployment of automated MRI quality control and AI-based image analysis systems.