深度學習用於頭頸癌淋巴結轉移的診斷與預後

Abstract

頭頸鱗狀細胞癌(Head and Neck Squamous Cell Carcinoma, HNSCC)是一種全球常見且具有高度侵襲性之惡性腫瘤，其治療困難度高且預後不佳。研究顯示，淋巴結轉移(Lymph Node Metastasis, LNM)乃是影響患者存活率及預後的重要因素。當前臨床上對淋巴結轉移的評估主要仰賴影像診斷及病理檢查，然而這些傳統評估方法存在許多臨床挑戰，包括診斷準確性不足、診斷過程具侵入性且風險較高、易受醫師主觀判斷影響，以及病理取樣存在明顯的空間限制和誤差問題。因此，急需一種非侵入性且高精確度的輔助診斷工具，以協助醫師進行更精準的臨床診療決策。 隨著人工智慧(Artificial Intelligence, AI)尤其是深度學習(Deep Learning, DL)技術的迅速崛起與發展，DL在醫學影像分析領域逐漸展現其優越的性能及強大的潛力 。深度學習透過多層次抽象特徵提取，能有效地學習複雜影像資料的潛在病理特徵，並顯著提升影像診斷與預後預測的準確性。因此，本研究提出一套基於深度學習之非侵入性診斷與預後預測框架，旨在克服現行傳統診斷方法的諸多限制，實現頭頸癌患者淋巴結轉移與治療效果的精確評估。 本研究設計分為兩個重要實驗階段，分別針對不同的影像資料與臨床需求進行深入的探索與驗證。 實驗A以194名經診斷確認的頭頸癌患者為研究對象，透過融合18F-FDG PET/CT影像建立無需影像分割的深度學習模型，利用先進的自監督學習(Self-supervised Learning)架構，包括Bootstrap Your Own Latent (BYOL)、Swin Transformer V2 (SwinV2)及Vision Transformer (ViT)等模型，以提升資料利用率並避免傳統影像分割產生的誤差 。為進一步強化模型泛化能力，實驗A導入資料增強策略，包括Mercator投影與螺旋變換，擴充訓練樣本，提升模型的預測表現 。最後，透過集成投票分類器，成功整合不同深度學習模型的特性，達到模型性能的最佳化。 實驗B則聚焦於162名接受放射治療之患者，利用初始診斷前後與治療期間的自適應放射治療(Adaptive Radiation Therapy, ART)模擬CT影像，透過創新的深度對比學習(Deep Contrastive Learning, DCL)方法，顯著提升影像特徵的辨識能力。同時，為避免單一模型過擬合問題，本研究亦提出了多模型集成框架，有效整合不同深度學習模型的預測結果，大幅提高了模型在臨床應用中的穩定性與準確性。 研究結果證實，本研究所提出的兩套深度學習框架在頭頸癌患者的預後評估中，表現均優於傳統臨床評估參數。 在實驗A中，利用PET/CT影像的集成模型在預測局部復發、淋巴結復發及遠端轉移方面，其ROC曲線下面積(AUC)分別達到0.850、0.878與0.893。 在實驗B中，融合基線與治療中CT影像的集成模型在預測相同臨床目標時，AUC則分別為0.773、0.747與0.793。 這些結果顯示，本研究開發的模型能有效提升臨床診斷與預後決策的準確性，特別是在利用PET/CT影像預測遠端轉移方面展現了卓越的性能。 綜合而言，本研究所開發之深度學習診斷與預後系統能在治療初期即精確地識別出高風險病患，提供更為精細且個性化的治療策略建議，有效改善患者的治療成效與長期生存機會。此外，本研究成果不僅提供了一個非侵入性、高準確性、且具泛化能力的臨床輔助工具，未來更可透過多中心的臨床驗證與推廣，期望將其廣泛應用於頭頸癌治療的精準醫療策略中，促進整體醫療資源的有效配置，推動頭頸癌治療朝向更具個人化與精確化的方向邁進。

Head and Neck Squamous Cell Carcinoma (HNSCC) is a globally prevalent malignancy known for its aggressive nature, treatment complexity, and poor prognosis. Among the critical factors affecting patient survival and prognosis, lymph node metastasis (LNM) has been recognized as a particularly significant determinant. Currently, clinical evaluation of lymph node metastasis primarily relies on imaging diagnostics and pathological examinations. However, these traditional methods present several clinical challenges, including insufficient accuracy, invasiveness, high procedural risk, susceptibility to subjective interpretation by clinicians, and inherent limitations due to sampling errors in biopsy. Consequently, there is an urgent need for non-invasive and highly accurate diagnostic tools to assist clinicians in making precise clinical decisions. The rapid advancement of Artificial Intelligence (AI), especially Deep Learning (DL) technologies, has introduced new possibilities in medical image analysis, demonstrating superior performance and considerable potential. Deep learning effectively extracts complex pathological features from medical images through multi-layer abstract feature extraction, significantly enhancing the accuracy of diagnostic assessments and prognostic predictions. Hence, this study proposes a non-invasive diagnostic and prognostic prediction framework based on deep learning, aiming to overcome existing limitations of conventional methods and accurately assess lymph node metastasis and treatment outcomes in HNSCC patients. This research is structured into two primary experimental phases, addressing diverse imaging datasets and clinical requirements comprehensively. Experiment A included 194 diagnosed HNSCC patients, utilizing a segmentation-free deep learning model developed through the fusion of 18F-FDG PET/CT images. Advanced self-supervised learning architectures, such as Bootstrap Your Own Latent (BYOL), Swin Transformer V2 (SwinV2), and Vision Transformer (ViT), were employed to enhance data utilization and circumvent segmentation-induced errors. To further augment model generalizability, Experiment A introduced sophisticated data augmentation techniques, including Mercator projection and spiral transformations, expanding training datasets and improving predictive performance. Ultimately, an ensemble voting classifier effectively integrated the unique strengths of various deep learning models, optimizing overall predictive performance. Experiment B focused on 162 patients undergoing radiation therapy, employing baseline diagnostic CT images and adaptive radiation therapy (ART) simulation CT scans. Through innovative Deep Contrastive Learning (DCL) methodologies, this experiment significantly enhanced the identification of critical image features. Additionally, to mitigate overfitting commonly observed in single-model architectures, a multi-model ensemble framework was introduced, substantially increasing model stability and predictive accuracy in clinical applications. The experimental results confirm that both deep learning frameworks proposed in this study significantly outperform traditional clinical assessment parameters in predicting outcomes for HNSCC patients. In Experiment A, the ensemble model utilizing baseline ¹⁸F-FDG-PET/CT images achieved an Area Under the Curve (AUC) of 0.850, 0.878, and 0.893 for predicting local recurrence, nodal relapse, and distant metastasis, respectively. In Experiment B, the merged ensemble model fusing baseline and ART simulation CT scans yielded AUCs of 0.773, 0.747, and 0.793 for the same clinical endpoints, with corresponding accuracies of 72.4%, 74.7%, and 75.8%. These findings demonstrate that the developed models effectively enhance the precision of clinical diagnosis and prognostic decision-making, with the PET/CT-based approach showing exceptional performance in predicting distant metastasis. In conclusion, the comprehensive deep learning diagnostic and prognostic system developed herein allows for early and precise identification of high-risk patients, facilitating more refined and personalized treatment strategies and significantly improving patient treatment outcomes and long-term survival quality. This study not only provides a non-invasive, high-accuracy, and highly generalizable clinical decision-support tool but also lays the groundwork for future multicenter clinical validation and widespread implementation in precision medicine strategies for HNSCC. Ultimately, this innovation aims to optimize the allocation of medical resources and propel HNSCC treatment towards greater personalization and accuracy.