基於腹部超音波影像之非酒精性肝病從脂肪肝至肝纖維化的階段性診斷

Abstract

非酒精性肝臟疾病（Non-Alcoholic Fatty Liver Disease, NAFLD）是一種常見的慢性肝病，約影響全球四分之一人口，主要特徵為肝細胞內異常脂質堆積，若未及時介入治療，可能進一步惡化為非酒精性脂肪性肝炎、肝纖維化、肝硬化，甚至發展為肝細胞癌。由於NAFLD屬於進展性疾病，因此在疾病早期階段如單純性脂肪肝與肝纖維化仍具可逆性，因此準確辨識這兩階段病程，對於臨床預後評估與治療介入具有關鍵意義。 目前臨床上廣泛採用 B 型超音波（B-mode ultrasound）作為非侵入性肝臟影像檢查方式，但其診斷表現易受到影像品質不佳、操作者經驗及主觀解讀影響，導致診斷準確性與一致性不足。為克服此限制，本研究提出一套基於深度學習電腦輔助診斷（Computer-Aided Diagnosis, CAD）架構，包含兩個針對不同任務所設計的模型，分別用於腹部超音波影像診斷NAFLD和肝纖維化，其中模型訓練階段以受控衰減參數（Controlled Attenuation Parameter, CAP, 單位：dB/m）與肝臟彈性係數（Elasticity, 單位：kPa）作為標註依據。 針對脂肪肝分類任務，本研究提出一套稱為Potent Boosts Channel-aware Separable ConvNeXt（PBCS-ConvNeXt）之深度學習架構，整合三項關鍵模組以強化分類能力：（1）Stem Cell Module，為一可訓練之高階影像預處理框架，用於強化基礎影像特徵；（2）Enhanced ConvNeXt Blocks，藉由加強通道間資訊的表徵能力，提升特徵抽取效率；（3）Boosting Block，透過多階段特徵融合，強化模型對於病灶區域的判別能力。 針對辨識難度較高的肝纖維化分類任務，導致辨識難度遠高於脂肪肝分類。因此，本研究提出CE-NeXt-CAT with text fusion架構。模型整合三項關鍵技術：（1）Convolution-of-Experts (CoE) 模組，透過專家路由處理異質特徵；（2）Channel-wise Adapter Tuning (CAT)，於凍結主架構下實現通道層級的高效微調；（3）Multimodal Text Fusion，將簡化語意資訊融合影像學習特徵以提升纖維化分類的能力，此設計特別適用於纖維化病灶表現模糊、形態差異顯著的分類情境，顯著提升模型整體辨識穩定性與臨床適應性。 實驗採交叉驗證，結果顯示PBCS-ConvNeXt模型於脂肪肝分類中達成約 82%的準確率與81%的靈敏度，接收者操作特徵曲線下面積（Area Under the Curve, AUC）達 0.88；CE-NeXt-CAT融合文字的模型於肝纖維化分類中則達成 76% 的準確率與 79% 的靈敏度，AUC 值達 0.80。整體而言，本研究提出之模型具備穩健性與臨床可行性，能於無需侵入式切片情況下實現NAFLD與肝纖維化的早期診斷，對疾病監控與長期治療策略具有實質效益。

Non-alcoholic fatty liver disease (NAFLD) is a prevalent chronic liver disorder that affects approximately one-quarter of the global population. The defining feature of NAFLD is abnormal lipid accumulation within hepatocytes. Without timely medical intervention, NAFLD may progress to more severe pathological conditions such as non-alcoholic steatohepatitis, liver fibrosis, cirrhosis, and even hepatocellular carcinoma. Given the progressive nature of NAFLD, early-stage conditions such as simple steatosis and liver fibrosis remain reversible. Accurate differentiation between these stages is crucial for prognosis and timely clinical intervention. B-mode ultrasound is commonly used for noninvasive liver assessment but is limited by poor image quality and dependence on the operator. To enable accurate classification, a deep learning-based computer-aided diagnosis (CAD) framework is proposed, comprising two task-specific models designed separately for NAFLD and liver fibrosis from a single abdominal ultrasound image. The controlled attenuation parameter (CAP, dB/m) and liver elasticity (kPa) serve as reference standards for annotation of NAFLD and fibrosis, respectively. For the NAFLD classification task, we develop an automatic model for fatty liver classification based on the ConvNeXt architecture, called potent boosts channel-aware separable intent-ConvNeXt (PBCS-ConvNeXt). The architecture of the study comprises three principal components: (1) the stem cell, a trainable framework engineered for advanced image preprocessing to establish a foundation for robust and diverse feature extraction; (2) enhanced ConvNeXt Blocks designed to amplify channel-wise features, thereby refining the processing capabilities; and (3) a boosting block that integrates stage-wise features to effectively extract and utilize information from ultrasound images. This tripartite structure enhances the ability of the model to identify fatty liver disease by optimizing the feature extraction and utilization from image data. For the fibrosis classification task, we extend the framework to CE-NeXt-CAT with text fusion, which integrates three architectural innovations: (1) a convolution-of-experts (CoE) mechanism for capturing heterogeneous texture patterns; (2) channel-wise adapter tuning (CAT) to support efficient fine-tuning across feature dimensions; and (3) a multimodal fusion strategy that incorporates simplified clinical descriptions, enabling cross-modal alignment between textual and imaging features. This multimodal architecture is particularly effective for capturing complex and ambiguous visual cues associated with liver fibrosis. Experimental results under 5-fold cross-validation demonstrate that the proposed diagnostic framework accurately identifies both NAFLD and liver fibrosis from abdominal ultrasound images. For NAFLD classification, the PBCS-ConvNeXt model achieved an accuracy of approximately 82%, a sensitivity of 81%, and a specificity of 83%. For liver fibrosis classification, the CE-NeXt-CAT with text fusion model attained an accuracy of 76%, a sensitivity of 79%, and a specificity of 80%. These findings highlight the robustness and clinical applicability of the proposed models for noninvasive liver disease assessment. By enabling reliable early detection of NAFLD and fibrosis without the need for biopsy, the framework supports timely intervention and improves long-term disease management strategies.