大腸直腸癌仍是全球癌症發生率與死亡率最高的癌症之一，而大腸鏡檢查的腺瘤偵測率（adenoma detection rate， ADR）則是目前最受認可且經充分驗證的品質指標。由於不同內視鏡醫師之間的腺瘤偵測率存在顯著差異，人工智慧輔助息肉偵測已成為一種自然的補充方式，前提是此類系統能夠適應真實臨床實務中的視覺異質性，包括不同的內視鏡設備、照明模式與病患族群。本論文以 REAL-COLON 全程大腸鏡影片基準資料集為研究對象，探討息肉偵測準確度與跨資料來源的泛化能力，而推論速度與前瞻性臨床部署則不在本研究的探討範圍內。首先，透過比較不同模型容量的偵測器，本研究重新界定了模型設計問題：相較於擴大偵測器本身的規模，將其訓練資料分布拓展至多種影像領域更具效益；以多元混合資料訓練、規模明顯較小的模型，無論在同領域或跨領域測試中，皆優於規模大得多、但僅以單一領域資料訓練的基準模型。此項觀察促成本論文的核心設計選擇：本研究不再進一步追求模型架構的規模擴張，而是採用單一、效能強大且已完成蒸餾的偵測器 RT-DETRv4-L，其訓練過程整合了 DINOv3 自監督特徵蒸餾，並將研究重點轉向擴展模型的學習來源，包括 REAL-COLON 資料、經頻率域適應處理的外部影格，以及針對模型自身失敗模式所設計的合成資料增強。此種領域擴展訓練策略產生了本論文的主要成果：相較於 RT-DETRv2-R50 參考基準模型，本方法在同領域資料上取得明顯改善，並在未曾見過的外部資料上獲得更大的相對提升，顯示其效益主要來自泛化能力的增強，而非對特定資料集的過度擬合。配套的消融實驗顯示，部分效能提升源自經領域適應處理的外部資料，且改善主要集中於先前未能偵測到的小型息肉；相較之下，針對模型弱點設計的合成資料增強所帶來的效益較小且一致性較低，其效果取決於訓練中已包含多少經適應處理的資料。本論文對此結果採取審慎解讀，認為關鍵在於訓練資料的多樣性，而非單純增加資料曝光量。本研究亦探討一種互補性的集成策略，將主要偵測器與一個輕量級協作網路結合；結果顯示，此策略以增加推論成本為代價換取有限的準確度提升，而不能被視為全面性的改進。它適用於優先考量準確度的情境，但無法取代基礎的訓練設計。由於先前針對此基準資料集發表的研究採用了不相容的逐影格評估方法，本論文僅與自行建立的基準模型進行比較，而不宣稱達到當前最佳水準。本論文亦針對基準偵測器進行失敗案例分析，結果顯示，其剩餘的召回率上限主要源自模型在常用操作解析度下系統性地漏檢小型息肉；此項診斷結果亦直接促成上述領域適應與資料增強策略的設計。未來研究方向包括採用更強大的自監督骨幹網路表徵、使用能在偵測器信心較高時維持嚴格定位準確度的信心條件式融合方法，以及透過前瞻性臨床驗證確認腺瘤偵測率的改善。;Colorectal cancer remains among the leading causes of cancer incidence and mortality worldwide， and the adenoma detection rate (ADR) of colonoscopy is its most validated qual ity indicator. Because this rate varies substantially across endoscopists， AI-assisted polyp de tection has emerged as a natural complement — provided such systems generalise across the visual heterogeneity of real clinical practice: different scopes， illumination modes， and patient populations. This thesis examines detection accuracy and cross-source generalisation on the REAL-COLON full-procedure benchmark， treating inference speed and prospective clinical de ployment as outside its scope. A first finding， obtained by contrasting detectors of different capacity， reframes the design problem: broadening a detector’s training distribution across imaging domains proves more valuable than enlarging the detector itself， with a substantially smaller model trained on a di verse mixture outperforming a much larger single-domain baseline both in- and out-of-domain. This observation motivates the thesis’s central design choice — rather than pursuing further ar chitectural scale， the work adopts a single strong， already-distilled detector (RT-DETRv4-L， whose training incorporates DINOv3 self-supervised feature distillation) and instead invests in extending what it learns from: REAL-COLON data combined with frequency-domain-adapted external frames and synthetic augmentation targeted at the model’s own failure modes. This domain-extended training strategy yields the thesis’s main result， a clear improve ment over the RT-DETRv2-R50 reference baseline in-domain， with an even larger relative gain on unseen external data — evidence that the benefit is driven by generalisation rather than by dataset-specific overfitting. An accompanying ablation attributes part of this gain to the domain adapted external data， concentrated specifically on previously-missed small polyps， while the synthetic weakness-augmentation contributes a smaller and less consistent effect that depends on how much adapted data is already present — a pattern the thesis interprets cautiously， as a question of training diversity rather than of raw exposure volume. A complementary ensemble strategy combining the main detector with a lightweight partner network is examined and found to trade additional inference cost for a modest accuracy gain， rather than constituting an outright improvement — useful in accuracy-prioritised settings but not a substitute for the underlying training design. Because prior published references on this benchmark use an incompatible per frame evaluation protocol， this thesis restricts its comparisons to its own baselines rather than claiming state-of-the-art status. The thesis also provides a failure analysis of baseline detectors， tracing their residual recall ceiling to systematic misses on small polyps at typical operating resolutions — a diagnosis that directly motivated the domain-adaptation and augmentation strategy above. Future directions include stronger self-supervised backbone representations， confidence-conditional fusion that preserves strict localisation accuracy when the detector is confident， and prospective clinical validation of adenoma-detection-rate improvement.
