應用自監督學習與主動學習於皮膚光學同調斷層掃描影像以提升標註效率

Abstract

深度學習已於醫療影像分析領域取得廣泛成就，於微米級皮膚光學同調斷層掃描（Optical Coherence Tomography; OCT）影像上亦不例外。然而，訓練高效能深度學習模型常仰賴大量人工標註，而OCT影像之標註須由醫師逐影像判讀，人力與時間成本高昂。面對此一挑戰，本研究系統性地探討三種可降低標註需求之策略，分別為資料增強、自監督預訓練與主動學習，並以七類別皮膚病灶分類任務量化其成效。實驗結果顯示，主動學習能有效選取高資訊量影像優先標註，將原先需 100%（2,032 張）標註量方能達到的 88.2% 分類準確率，降低至僅需約 50%。若進一步配合 SimCLR 自監督預訓練與四倍影像翻轉增強，最終僅需約 25% 至 30% 的標註量即可達標，三者合用可將所需標註影像縮減至約 500 至 600 張，節省逾 1,400 張影像標註工作量。針對主動學習，本研究採用基於不確定性、基於多樣性與混合型共七種演算法進行全面比較，發現各方法所能帶來之效益接近。其中，我們觀察到「類別重分布現象」：主動學習傾向將標註預算重新分配予難以正確分類之少數病灶類別，如曬斑與脂漏性角化症，針對此類病灶最高可分別提升 35.3% 與 29.4% 之召回率。透過質性比較各演算法所選取的影像，本研究進一步歸納其病理特徵差異，以闡明主動學習於皮膚OCT影像上的運作機制。舉例而言，基於不確定性的演算法傾向優先挑選病灶特徵不明確之影像；基於多樣性的方法則傾向涵蓋不同對比度、噪點等影像風格多元之樣本。最後，本研究將上述策略延伸至細胞核語意分割任務，發現自監督預訓練與主動學習在小標註量時能帶來一定效能提升，然隨標註量增加其效益趨於消退；資料增強策略則在各標註量下皆具穩定效益。本研究期望所呈現之成果能提供實務參考，鼓勵OCT影像相關研究與臨床應用積極導入此類策略以提升標註效率，並進一步推廣至其他標註成本高昂的醫學影像任務。

Deep learning has achieved widespread success in medical image analysis, including its application to cellular-resolution skin Optical Coherence Tomography (OCT). However, training high-performance models typically requires large volumes of annotated data. Since annotating OCT images demands per-image review by expert physicians, it incurs substantial time and labor costs. To address this, this thesis systematically investigates three strategies for reducing annotation requirements: data augmentation, self-supervised pretraining, and active learning, quantifying their effectiveness on a seven-class skin lesion classification task. By prioritizing high-information images for annotation, active learning reduces the labeled data needed to reach an 88.2% classification accuracy from 100% (2,032 images) to around 50%. When combined with SimCLR self-supervised pretraining and four-fold image flipping, only 25–30% of the data needs to be labeled to attain the same target, lowering the requirement to roughly 500–600 images and saving over 1,400 annotations. For active learning, we compare seven algorithms spanning uncertainty-based, diversity-based, and hybrid approaches, observing a consistent class redistribution phenomenon in which the annotation budget is reallocated toward minority lesion classes that the model struggles to classify correctly (e.g., Solar Lentigo and Seborrheic Keratosis). Qualitative comparison of the selected images further reveals distinct pathological selection patterns: uncertainty-based methods tend to prioritize images with ambiguous lesion features, whereas diversity-based methods favor samples with varied imaging styles, such as differing contrast and noise levels. Finally, we extend these strategies to a nuclei semantic segmentation task, finding that self-supervised pretraining and active learning yield moderate gains mainly under low annotation budgets but diminish as more labeled data becomes available, whereas data augmentation provides consistent improvements across all annotation levels. We hope these findings serve as a practical reference and encourage the adoption of active learning strategies in future OCT research and clinical workflows.