| 摘要: | 2024 年台灣衛生福利部統計處統計，胰臟癌已被列入 113 年十大癌症死因中的第七名，更是令人聞之色變的癌症。本研究針對胰臟在醫學影像分割任務中準確率偏低的情況，提出以深度學習技術提升胰臟自動分割的效能。胰臟因面積小、形狀變異大及邊界模糊，常常導致傳統分割方法難以達到高準確率。隨著電腦斷層掃描（CT）與人工智慧技術的發展，結合影像增強與深度學習模型已成為提升分割表現的重要方向。因此，本研究以 Transformer 架構為基礎，改良 SAM2 模型，並針對無顯影劑之 CT 影像當成資料集，並進行適當的前處理與特徵強化，期望提升胰臟及其他器官的分割準確率與穩定性。
在實驗設計上，本研究蒐集多個公開腹部 CT 資料集，並針對影像進行 HU 值調整、CLAHE 對比度增強等前處理，將每層影像視為影片幀依序的輸入模型。模型架構以改良後的 SAM2 為核心，並探討不同特徵融合策略、prompt encoder 與 memory encoder 設計對分割效能的影響。實驗結果顯示，本研究的方法於 NIH TCIA 資料集的胰臟分割 Dice Score達到 92.3%，優於現有方法，且在肝臟、腎臟、脾臟等多器官分割任務亦具不錯的泛化能力。最後進一步以 Grad-CAM 等技術分析模型特徵學習的狀況，驗證其於小器官辨識的解釋性，證明本研究方法在醫學影像分割領域的可行性與潛力。;According to statistics from the Ministry of Health and Welfare of Taiwan in 2024, pancreatic cancer ranked seventh among the top ten causes of cancer death, making it a particularly concerning disease. This study addresses the challenge of low segmentation accuracy for the pancreas in medical imaging by proposing a deep learning-based approach to improve automatic segmentation of the pancreas and multiple organs. Due to its small size, variable shape, and indistinct boundaries, the pancreas is difficult to segment accurately using traditional methods. With the advancement of computed tomography (CT) and artificial intelligence technologies, combining image enhancement with deep learning models has become an important direction for improving segmentation performance. Therefore, this research is based on a Transformer architecture, modifies the SAM2 model, and applies preprocessing and feature enhancement to non-contrast CT images, aiming to improve the segmentation accuracy and stability for the pancreas and other organs.
In the experimental design, multiple publicly available abdominal CT datasets were collected. Preprocessing steps such as HU value adjustment and CLAHE contrast enhancement were applied, and each CT slice was treated as a video frame for model input. The core model is a modified SAM2 architecture, and the effects of different feature fusion strategies, prompt encoder, and memory encoder designs on segmentation performance were explored. Experimental results show that the proposed method achieved a Dice Score of 92.3\% for pancreas segmentation on the NIH TCIA dataset, outperforming existing methods. The model also demonstrated good generalization in multi-organ segmentation tasks, such as for the liver, kidneys, and spleen. Further analysis using Grad-CAM and related techniques verified the model’s interpretability in recognizing small organs, confirming the feasibility and potential of the proposed approach in the field of medical image segmentation. |
