預測原發性醛固酮症腎上腺切除術後之結果：臨床生化特徵之機器學習模型暨腎上腺電腦斷層影像分割

Abstract

原發性醛固酮症是一種內分泌疾病，常見症狀為高血壓與低血鉀，主要分為醛固酮分泌腺瘤及雙側腎上腺增生等兩種亞型。單側腎上腺切除手術是治療醛固酮分泌腺瘤之首選，然而，儘管大多數患者在術後低血鉀症狀獲得改善，達到生化面向之完全成功，平均卻只有 37% 的患者能達到臨床面向之完全成功，亦即術後不必服用降血壓藥物，便可使血壓維持在正常標準下。 本研究以預測原發性醛固酮症患者經過腎上腺切除術後之結果為主軸，分為兩部分。第一部分利用 640 位生化完全成功之患者的術前臨床生化特徵，以機器學習模型預測術後的臨床結果。結果顯示，在 7 種機器學習模型中，隨機森林有最佳的分類表現，其 AUC 在 5 次分層 5 折交叉驗證上可達到 0.750±0.009 。第二部分則結合圖像神經網路與 U-Net 架構，設計出一深度學習影像分割模型，從 190 位原發性醛固酮症患者的腹部靜脈相電腦斷層細切影像中，劃分出腎上腺的確切區域。經由 4 次 5 折交叉驗證，本研究所提出之方法用於腎上腺分割的 DSC 為 0.800±0.002 。 本研究比較了不同前處理方式與機器學習分類模型的表現，證實隨機森林在預測術後臨床結果的潛力，並提出了有效的腎上腺影像分割方法，提高自動化分析腎上腺的可能性。期望未來能結合影像與臨床生化數據，對原發性醛固酮症患者的術後結果進行更加全面而準確的預測，將有助於臨床醫師根據個別患者的術前資料，制定更精確的治療計畫。

Primary aldosteronism (PA) is an endocrine disorder characterized by hypertension and hypokalemia. It is classified into two main subtypes: aldosterone-producing adenoma and bilateral adrenal hyperplasia. Unilateral adrenalectomy is the preferred treatment for aldosterone-producing adenoma. While most patients achieve complete biochemical success (resolution of hypokalemia) after surgery, only 37% on average achieve complete clinical success, defined as maintaining normal blood pressure without antihypertensive medications. This study investigates the prediction of postoperative outcomes for PA patients undergoing adrenalectomy, comprising two main components. The first component utilizes preoperative clinical and biochemical features from 640 patients who achieved complete biochemical success to develop machine learning models for predicting clinical outcomes. Among seven machine learning models, random forest demonstrated the best performance, achieving an AUC of 0.750±0.009 in 5 times repeated 5-fold stratified cross-validation. The second component introduces a deep learning segmentation model combining Graph Neural Networks (GNN) with a U-Net architecture to delineate adrenal regions from abdominal venous-phase CT images of 190 PA patients. The proposed method achieved a Dice Similarity Coefficient (DSC) of 0.800±0.002 through 4 times repeated 5-fold cross-validation. This study confirms the potential of random forest for predicting postoperative clinical outcomes and presents an effective adrenal segmentation approach that enhances the efficiency of automated adrenal analysis. Future work will aim to integrate imaging and clinico-biochemical data for more comprehensive and precise outcome predictions, providing valuable guidance for clinicians in tailoring treatment plans for PA patients.