融合多相位 CT 影像特徵及增強對齊技術之肝臟與腫瘤分割

Abstract

本研究旨在針對多相位電腦斷層（CT）影像中標註不全（incomplete labeling）所帶來的實務挑戰，提出並驗證一套結合影像對齊與模型融合的腫瘤分割策略。多相位影像具備豐富的對比資訊，為腫瘤分割帶來潛在效能提升空間，因此被視為提升腫瘤分割效能的重要方向。然而臨床上常僅對部分相位進行標註，使得無法直接應用於監督式訓練與有效融合。 為此，本研究以提升肝臟腫瘤切割準確度為目標，設計一系列跨相位融合實驗。首先，針對動脈相與靜脈相進行 Z 軸對齊與切片內配準，結合 Greedy Local 與 Bipartite Matching 等策略，搭配 SIFT 與 SURF 特徵進行配對，並引入肝臟區域限制與網格濾波以提升配準品質。透過此流程，我們嘗試將動脈相標註轉移至靜脈相，建構偽標註以進行弱監督訓練。 實驗結果顯示，動脈相模型在完整標註下可達 0.4762 之 DICE 分數，靜脈相模型雖受限於偽標註品質，表現較弱（DICE 約為 0.2），但透過決策層融合（decision-level fusion），結合兩相位模型後整體效能提升至 0.5627，展現弱信號融合之潛力。此外，輸入層融合（input-level fusion）進一步驗證配準的實質效果：未配準的輸入導致顯著效能下降（DICE 0.1807），而配準後之雙相位輸入則可達 0.5453，顯示幾何對齊為跨相位資訊整合的重要前提。 綜合而言，本研究不僅證實多相位融合可在標註不全下提升腫瘤分割效能，也系統性分析不同配準與融合策略之成效，提供日後發展弱監督式多相位醫學影像模型之實證基礎與實作建議。

This study addresses a practical challenge in multi-phase computed tomography (CT) imaging: the issue of incomplete labeling across phases. Although multi-phase imaging provides rich temporal contrast information that is potentially beneficial for tumor segmentation, clinical datasets often include annotations for only a subset of phases. This limitation restricts the feasibility of supervised training and reduces the effectiveness of cross-phase integration. To overcome this limitation, we designed a series of experiments aimed at improving liver tumor segmentation by integrating image registration and model fusion techniques. Specifically, we performed Z-axis alignment and in-plane registration between the arterial phase (C+A) and the venous phase (C+P), employing pairing strategies such as Greedy Local Matching and Bipartite Matching, along with SIFT and SURF features for keypoint matching. Additional constraints such as liver-region masking and grid-based filtering were applied to enhance registration quality. Based on the resulting transformations, we transferred annotations from the C+A phase to the C+P phase to generate pseudo-labels for weakly supervised training. Experimental results showed that the C+A model trained with complete annotations achieved a DICE score of 0.4762. Although the C+P model trained using pseudo-labels performed poorly with a DICE score around 0.2, decision-level fusion of the C+A and C+P models improved segmentation performance to 0.5627. Furthermore, input-level fusion experiments demonstrated the importance of spatial alignment. Without registration, fused inputs achieved only 0.1807 in DICE score, whereas alignment raised the score to 0.5453. These results confirm that while alignment alone does not resolve semantic differences between phases, it is essential for effective multi-phase integration. In summary, this study demonstrates the feasibility of multi-phase fusion under incomplete labeling and provides a systematic evaluation of registration and fusion strategies. The proposed framework offers empirical insights and methodological guidance for advancing weakly supervised tumor segmentation in real-world clinical imaging scenarios.