使用集成學習與多相位腹部電腦斷層影像預測肝細胞癌之BCLC分期系統

Abstract

肝癌為我國十大癌症發生率第五位及死亡率第二位，是重要的醫療問題。肝癌的早期預測可透過電腦斷層影像進行診斷。在注入顯影劑後，腫瘤會明顯呈現於影像中，且會於短時間內照攝多張影像以觀察不同相位腫瘤的型態。醫師會根據腫瘤狀況，初步診斷患有惡性肝細胞癌之患者，並使用近年來臨床醫師常用的分期系統進行分期，以對應不同治療方式，如巴塞隆納臨床肝癌分期(Barcelona Clinic Liver Cancer classification, BCLC)系統。上述流程皆依照醫師經驗進行診斷，缺乏自動性及方便性，診斷過程花費時間，為臨床醫師的一大痛點。為解決上述問題，本研究建立多個模型來改善現階段的診斷過程。主要目標為實現多相位影像融合，包含三種深度學習方法：early fusion、intermediate fusion、late fusion及一種影像處理方法。前處理階段包含使用nnU-Net學習影像中肝臟及肝腫瘤特徵，接著保留肝臟區域以排除其他腹部器官的干擾，再進行不同相位間的融合與分類，最後使用了集成學習來集成不同模型以精準分期，並提供腫瘤大小與顆資訊。最終模型之準確度、精準度、敏感度及特異度分別約為78%、87%、57%及76%；三期之準確度為93.33%、73.33%,及66.67%；腫瘤顆數與大小之平均誤差為0.3與0.4。本研究可有效輔助醫師診斷並加速診療過程，自動化分析也將解決臨床醫師的痛點。

Liver cancer poses a significant healthcare challenge, ranking as the fifth most common cancer and the second leading cause of death among individuals. Early detection of liver cancer is crucial for successful treatment, and computed tomography (CT) imaging plays a vital role in the diagnosis process. Through administering a contrast agent, tumors are rendered visible in CT scans. These scans are captured in successive phases, enabling a detailed examination of the tumor's morphology. Physicians use these images and staging systems, such as the Barcelona Clinic Liver Cancer (BCLC) classification, to make preliminary diagnoses of hepatocellular carcinoma (HCC) and determine appropriate treatment strategies. However, the current diagnostic process heavily relies on physician expertise, lacks automation, and can be time-consuming, posing challenges for clinical practitioners. To overcome these challenges, this study proposed using deep learning models to enhance the diagnostic process. The primary objective is to achieve multi-phase image fusion, involving three deep learning approaches: early fusion, intermediate fusion, and late fusion, combined with an image processing method. In the preprocessing stage, we employ nnU-Net to extract essential liver and tumor features from the images. The liver region is subsequently isolated to remove interference from other abdominal organs. Following this, fusion and classification among different phases are conducted. Finally, ensemble learning is employed to integrate diverse models for accurate staging, as well as to provide valuable tumor size and count information. Finally, ensemble learning techniques were employed to integrate different models to classify BCLC stages and provide tumor size and count information. The overall accuracy of the final model was approximately 78%, with precision, sensitivity, and specificity values of 87%, 57%, and 76%, respectively. The accuracy for stages A, B, and C was 93.33%, 73.33%, and 66.67%, respectively. The average tumor count and size errors were 0.3 and 0.4, respectively. The proposed method effectively assists physicians in diagnosing, accelerates the diagnostic process, and automates analysis, addressing clinical practitioners' challenges.