以nnU-Net框架進行遷移式學習運用於犬貓腹腔斷層多器官自動分割

Abstract

隨著獸醫影像診斷技術的進步，電腦斷層掃描儀器的普及，已逐漸成為評估犬貓腹腔疾病的黃金標準之一。然而，傳統的人工影像分割方式，不僅耗時又費力，且具有高度的主觀性，成為限制臨床量化分析的瓶頸。此外，獸醫領域面臨標注數據數量的稀缺，以及動物體之間，如犬貓品種間，巨大的解剖變異挑戰。本研究旨在建立一套以nnU-Net為核心框架，並結合遷移式學習策略的自動化多器官分割系統。本研究使用24例臨床犬貓腹腔顯影增強電腦斷層影像數據，並對比了兩種訓練配置：(1)套用大規模人類電腦斷層影像數據資料集預訓練權重之較低解析度遷移學習模型；(2)以隨機初始化進行高解析度之從頭訓練模型。研究結果顯示，遷移式學習模型展現了卓越的效能與穩定性，首先在分割精確度上，遷移學習模型整體在測試資料集中的平均Dice相似係數達到0.931，而在同樣訓練週期下，遷移學習模型能有效緩解小樣本數據的過擬合問題。其次是收斂效率，遷移學習僅需約6個Epochs，即可使Dice相似係數突破0.8，而收斂速度較從頭訓練快約95%。再者是推論效率與技術發現，遷移學習模型推論單一病例僅需114秒，而從頭訓練模型則需高達2049秒。經分析數據規劃發現，此速度差距源於預處理的邏輯差異，從頭訓練模式為捕捉細微結構，自動配置了極細的體素間距，導致處理的體素量高達222.82百萬，而遷移學習則繼承了人類的配置設定，處理量僅需12.32百萬。最後是跨物種泛化，遷移學習模型在貓的測試案例中表現穩定，但從頭開始訓練模型當過度訓練則會出現嚴重失效。最終本研究證實，遷移式學習不僅能顯著提升獸醫影像在人工智慧的準確度，更透過優化的計算配置，解決了臨床部署的效率負擔。遷移式學習僅需不到兩分鐘的推論時間，使其具備與醫院影像系統整合的高度潛力，以提升精準獸醫醫學關鍵技術的支持。

With advances in veterinary diagnostic imaging, computed tomography (CT) has become a gold standard for evaluating abdominal diseases in dogs and cats. However, manual image segmentation is time-consuming and highly subjective, posing a major bottleneck for quantitative clinical analysis. In addition, the veterinary domain faces challenges from data scarcity and substantial anatomical variability across breeds. The objective of this study was to develop an automated multi-organ segmentation system based on the nnU-Net framework using transfer learning. Utilizing a clinical dataset of 24 contrast-enhanced canine and feline abdominal CT cases for training, this study compared two configurations: (1) a lower-resolution transfer learning model initialized with pretrained human CT weights, and (2) a high-resolution training-from-scratch model based on random initialization. The results demonstrate that the transfer learning model achieved superior performance. First, it attained an overall mean Dice Similarity Coefficient (DSC) of 0.931 on the test set, effectively mitigating overfitting under identical training durations. Second, the transfer learning model surpassed a DSC of 0.8 within six epochs, converging nearly 95% faster than Training from Scratch. Third, the transfer learning model required only 114 seconds per case, whereas the Training from Scratch model took 2,049 seconds. This discrepancy stemmed from preprocessing configurations: the Training from Scratch mode automatically adopted fine voxel spacing, processing 222.82 million voxels, while Transfer learning processed only 12.32 million voxels. Finally, the Transfer learning model maintained stable performance on feline cases, while the Training from Scratch model exhibited severe degradation when overtrained. In conclusion, transfer learning significantly improves segmentation accuracy and addresses the computational burden for clinical deployment. With an inference time of under two minutes, this approach demonstrates strong potential for integration into veterinary hospital imaging systems.