基於群集的潛在增強通過元學習對病理全切片影像分類;Cluster-based Latent Augmentation via Meta-Learning for Whole Slide Image Classification

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/93368

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/93368

题名:	基於群集的潛在增強通過元學習對病理全切片影像分類;Cluster-based Latent Augmentation via Meta-Learning for Whole Slide Image Classification
作者:	潘祐家;Wicaksana, Prabowo Yoga
贡献者:	資訊工程學系
关键词:	病理全切片影像分析;元學習;群集分析;潛在空間增強;集成分類器;Whole slide images;Meta-learning;Clustering;Latent space augmentation;Ensemble classifier
日期:	2023-07-28
上传时间:	2024-09-19 16:56:32 (UTC+8)
出版者:	國立中央大學
摘要:	病理全切片影像（WSI）分類涉及分析以超高解析度掃描整個玻片捕獲的生物樣本或組織樣本的大型數位影像。 WSI 分類的目標是準確地識別和分類圖像中不同的感興趣區域，例如癌細胞或正常組織以及特定的細胞類型和結構。 WSI 分類在數字病理學和醫學研究領域中具有眾多應用，包括幫助診斷和治療癌症，識別潛在的藥物標靶，以及實現個體化治療。近年來，深度學習技術的最新進展已經在 WSI 分類性能方面取得了顯著的改進，最先進的模型實現了高水準的準確性和穩健性。然而，當前的 WSI 分類管道通常需要大量標記數據進行訓練，獲取這些數據既昂貴又耗時。此外，簡單地在大型數據集上訓練模型並不能保證良好的泛化性能。事實上，由於對訓練數據的過度擬合和難以學習魯棒特徵等問題，WSI 模型在大型數據集上訓練時的泛化能力往往受到限制。在這項工作中，我們提出了一個框架來處理上述問題。我們將集成分類器方法與元學習相結合，使模型能夠從有限數量的標記樣本中學習，同時仍能取得卓越的性能。此外，我們還提出了一種簡單的基於集群的噪聲注入，以強制模型學習更穩健和多樣化的特徵。通過綜合實驗，我們僅在兩個公開可用的 WSI 分類任務數據集上使用少量樣本證明了我們方法的有效性。;Whole Slide Image (WSI) classification involves analyzing large digital images of tissue samples or other biological specimens captured by scanning the entire slide at ultra resolution. The objective of WSI classification is to accurately identify and classify different regions of interest within the image, such as cancerous or normal tissue, as well as specific cell types and structures. WSI classification has numerous applications in the field of digital pathology and medical research, including aiding in the diagnosis and treatment of cancer, identifying potential drug targets, and enabling personalized medicine. Recent advances in deep learning techniques have led to significant improvements in WSI classification performance, with state-of-the-art models achieving high levels of accuracy and robustness. However, current WSI classification pipelines typically require a large amount of labeled data for training, which can be both costly and time-consuming to obtain. Moreover, simply training a model on a large dataset does not guarantee good generalization performance. In fact, the generalization ability of WSI models is often limited when trained on large datasets, due to issues such as overfitting to the training data and difficulty in learning robust features. In this work, we propose a framework to deal with the aforementioned issues. We incorporate ensemble classifier approach integrated with meta-learning, which enables the model to learn from a limited number of labeled samples while still achieving remarkable performance. Furthermore, we also propose a simple cluster-based noise injection to force the model to learn more robust and diverse features. Through comprehensive experiments, we demonstrate the effectiveness of our approach with only a small number of samples on two publicly available datasets for WSI classification tasks.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	14	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....