體素層級監督式對比學習融合 3D U-Net 之冠狀動脈光學同調斷層影像鈣化斑塊分割

Abstract

冠狀動脈鈣化（Coronary Artery Calcium, CAC）與不良臨床預後相關，亦可能影響經皮冠狀動脉介入治療（Percutaneous Coronary Intervention, PCI）之處置策略與手術結果。臨床上，對其範圍與嚴重程度之辨識與量化具有重要價值。血管內光學同調斷層掃描（Intravascular Optical Coherence Tomography, IVOCT）可提供高解析度影像導引，然而單次 Pullback 之影像量龐大，且逐幀人工標注成本高，使得即時且一致的全面評估難以達成。因此，冠狀動脉斑塊的自動化分析方法具有重要的臨床應用價值。 本研究旨在建立冠狀動脈OCT影像之鈣化斑塊自動分割方法，以降低逐幀人工標注負擔，並支援後續臨床量化分析與治療規劃。研究資料依病患爲單位分為訓練、驗證與測試資料集，訓練集包含 125 名病患，驗證集與測試集各 6 名病患；對應影像幀數分別為 31,967、1,499 與 1,482幀。 方法上，本研究以監督式學習之三維 U 形網路（3D U-Net）作爲分割骨幹，引入了體素層級的對比式學習（Voxel-level Contrastive Learning），用於在冠狀動脉光學同調斷層掃描（OCT）影像中分割冠狀動脉鈣化。針對鈣化病灶呈小體積或薄片狀與類別嚴重不平衡所導致之漏檢與邊界不穩定問題，該對比式分支以標注爲依據定義正負樣本關係，直接在特徵空間强化鈣化與背景之可分離性，以期實現局部細緻分割，並降低漏檢風險。此外，本研究引入類別導向之伫列式記憶庫以累積跨批次特徵，藉此擴充可用樣本對並降低小批次取樣偏差，使樣本集合可隨訓練歷程持續更新。另爲避免負樣本集合被大量且易區分的背景所主導，進而稀釋對比學習的有效訊號，本研究採用困難負樣本挖掘結合隨機抽樣的混合策略，以實現聚焦於易混淆背景的同時兼顧背景多樣性之維持。 結果顯示，體素層級的對比式學習結合3D U-Net的方式在冠狀動脈OCT影像之鈣化斑塊分割中呈現良好效能表現，相較於3D U-Net基礎模型，本研究方法取得穩定但幅度有限之提升，達到 Dice 係數（Dice Coefficient, DSC）0.8383±0.0127 ，交並比（Intersection-over-union, IoU）0.7216±0.0109，顯示鈣化斑塊預測區域與人工標注之空間重叠度良好。精確率（Precision）0.8825±0.0134、平衡準確率（Balanced Accuracy, BalAcc）0.8964±0.0136 與 馬修斯相關係數（Matthews Correlation Coefficient，MCC）爲0.8313±0.0126。值得注意的是，靈敏度達到（Sensitivity；亦即 Recall） 0.7984±0.0121，比之3D U-Net基礎模型上升了1.99%，體現其增加了鈣化之檢出。影像層面之觀察亦指出，本方法對體積較小或薄片狀鈣化病灶能提供較穩定之分割邊界。 綜合而言，本研究提出之體素層級監督式對比式學習，透過標注導向的表徵學習機制，提升小目標且嚴重類別不平衡情境下之鈣化分割能力，具有作爲冠狀動脉鈣化量化流程輔助工具之潜力。

Coronary artery calcification (CAC) is associated with adverse clinical outcomes and may influence both procedural strategy and outcomes in percutaneous coronary intervention (PCI). Accurate identification and quantification of CAC extent and severity are therefore clinically important. Intravascular optical coherence tomography (IVOCT) provides high-resolution imaging guidance; however, a single Pullback contains a large number of frames, and frame-by-frame manual annotation is labor-intensive. These factors make real-time, consistent, and comprehensive assessment difficult in routine practice. Consequently, automated analysis of coronary plaque has substantial clinical value. This study aims to develop an automatic calcified plaque segmentation method for coronary OCT images to reduce the burden of frame-level manual annotation and to support downstream quantitative analysis and treatment planning. The dataset was split at the patient level into training, validation, and test sets. The training set included 125 patients, while the validation and test sets each included 6 patients, corresponding to 31,967, 1,499, and 1,482 frames, respectively. Methodologically, we used a 3D U-Net as the supervised segmentation backbone and incorporated voxel-level contrastive learning to segment calcified plaque in coronary OCT images. To address missed detections and unstable boundaries caused by small-volume or thin, sheet-like lesions and severe class imbalance, the contrastive branch defined positive and negative pairs using manual annotations. This design explicitly increases the separability between calcification and background in feature space, with the goal of improving local, fine-grained segmentation and reducing false negatives. In addition, we introduced a queue-based memory bank with class-aware sampling to accumulate cross-batch embeddings, thereby expanding the pool of available contrastive pairs and alleviating mini-batch sampling bias. To avoid the negative pool being overwhelmed by abundant, easily separable background, which would weaken the contrastive signal, we further adopted a hybrid strategy that combines hard negative mining with random sampling. This approach emphasizes confusing background regions while maintaining background diversity. The proposed method demonstrated strong performance for calcified plaque segmentation in coronary OCT images. Compared with the 3D U-Net base model, it achieved a consistent but modest improvement, reaching a Dice coefficient (DSC) of 0.8383 ± 0.0127 and an intersection-over-union (IoU) of 0.7216±0.0109, indicating good spatial overlap with manual annotations. It also achieved a precision of 0.8825 ± 0.0134, a balanced accuracy (BalAcc) of 0.8964 ± 0.0136, and a Matthews correlation coefficient (MCC) of 0.8313 ± 0.0126. Notably, sensitivity (Recall) reached 0.7984 ± 0.0121, representing a 1.99% increase over the 3D U-Net baseline and suggesting improved detection of calcified lesions. Qualitative inspection further indicated that the proposed method produced more stable boundaries for small or thin, sheet-like calcifications. In summary, we propose a voxel-level supervised contrastive learning framework that leverages annotation-guided representation learning to improve calcification segmentation under small-target and severely imbalanced conditions. The method shows potential as an assistive tool for coronary calcification quantification workflows.