時間： 112年12月11日(星期一) 19:00-21:00

地點：大仁樓301室

主持人：左瑞麟 老師

演講者：聯發科技 黃啟賢博士

演講題目：自督導式結構學習之內視鏡影像息肉分類與加密方法和隱私保護

Classification of Polyps in Endoscopic Images using Self-Supervised Structured Learning and its Encryption and Privacy Preservation

演講摘要：本研究提出了一個新穎的兩階段電腦輔助診斷 (CAD) 學習方式，採用具有自督導式學習 (SSL) 的捲積神經網路 (CNN) 將息肉準確分類為增生性息肉 (HP) 或管狀腺瘤 (TA) ). 我們使用SimCLR 中的對比學習（Contrastive learning）與look-into-object (LIO)來關注整體息肉區域，從而提高了模型性能。由於醫學影像很難獲取，沒有足夠的醫學影像來訓練息肉分類的有效表示，因此另一種方法使用自然圖像來取代息肉圖像來完成預訓練任務。有鑑於使用深度神經網路 (DNN) 進行醫學診斷的隱私問題，先前的研究提出了可學習圖像加密的概念。雖然一些方法已經部分攻擊了以前的加密方式，但仍有改進的空間。為了解決隱私問題，我們也提出了一種增強的可學習圖像加密方式，不僅可以訓練強大的DNN模型，還可以確保訓練圖像的隱私。研究實驗結果，我們提出自督導式學習的一個強健性模型，可增強息肉結構資訊和有效分類，該模型能夠透過遷移式學習準確地將息肉分類為 HP 或 TA。透過有效利用少量標記的息肉圖像，使用 ResNet-18 的模型學習息肉結構特徵表示。在準確性和 F1-score衡量指標優於現有方法。 此外，加密結果證明了我們的方法在實現高性能同時保護隱私方面的有效性。

This study introduces a novel two-stage learning computer-aided diagnosis (CAD) scheme that employs a convolutional neural network (CNN) with self-supervised learning (SSL) for accurate classification of polyps into hyperplastic polyps (HP) or Tubular Adenomas (TA). Our proposed model incorporates look-into-object (LIO) and contrastive learning in SimCLR to emphasize the comprehensive polyp region, resulting in improved model performance. To overcome the challenge of limited medical images for efficient polyp representation, an alternative approach utilizing natural images for the pretext task is explored. In light of the privacy concerns surrounding medical diagnoses using deep neural networks (DNNs), previous studies have proposed the concept of learnable image encryption. While some methods have partially attacked previous encryption schemes, there is still room for improvement. To address this, we propose an enhanced learnable image encryption scheme that not only trains a powerful DNN model but also ensures the privacy of training images. The experimental results validate the effectiveness of our proposed scheme, which leverages polyp object structure information and self-supervised learning to produce a robust model capable of accurately classifying polyps as HP or TA in the prediction head through backbone transfer. By effectively utilizing limited labeled polyp images, the backbone model, based on ResNet-18, focuses on the holistic polyp representation. The proposed scheme outperforms existing methods in terms of accuracy and F1-score. Furthermore, the encryption results demonstrate the efficacy of our method in achieving high performance while preserving privacy.