卷積神經網路訓練樣本量對肝細胞癌病理影像分類的影響

Abstract

肝細胞癌（Hepatocellular carcinoma）是一種原發性肝臟惡性腫瘤，每年導致全球數十萬人死亡。早期的病理學檢查依靠病理科醫師使用顯微鏡觀察細胞與組織型態進行診斷。隨著科學技術的進步，病理學檢查加入了人工智慧輔助醫生診斷，讓診斷比以往更快、更準確。然而，數位病理切片影像的解析度往往超過數十億像素，因此需要將其切割成數以萬計的補丁（patch）才能應用。實驗室過去的研究主要專注在數位病理切片影像的切片數量如何提高模型對新數位病理切片影像分類的正確率，但沒有探討數位病理切片影像中的補丁數量如何影響模型。本研究探討不同的補丁訓練樣本量對卷積神經網絡模型診斷正確率的影響，希望可以減輕病理科醫師在標註數位病理切片影像工作的負擔。實驗結果顯示可以用少量的醫師標註補丁達成該數位病理切片影像之分類模型訓練並取得相當好的分類正確率（多高於90%以上）。此外，本研究還發現雖然添加數位病理切片影像的數量可以提高模型對新數位病理切片影像進行分類的正確率，但從單張數位病理切片影像中選取更多的補丁進行訓練，並不能有效提高新數位病理切片影像分類的正確率。本論文提供了增進數位病理切片影像分類正確率的方向，同時可以減少病理科醫師於標註資料的負擔，降低分類器訓練之計算時間與成本。

Hepatocellular carcinoma is a primary liver malignancy and causes hundreds of thousands of deaths worldwide each year. Early pathological examination relied on pathologists using microscopes to observe the morphology of cells and tissues to diagnose. As technology advances, the pathological examination has added artificial intelligence to assist pathologists in making diagnoses faster and more accurate. However, digital pathological images often have resolutions in billions of pixels, so they need to be cropped into tens of thousands of patches for application. Our past research has mainly focused on how the number of images improves the model's accuracy for new images but has not explored how the number of patches in an image affects the model's accuracy. This study explores the effect of different training patch sample sizes on the accuracy of convolutional neural network models, hoping to reduce the burden of pathologists on labeling images. The experimental results show that a small number of labeled patches can be used to implement the training of a pathological image classification model, and good classification accuracy (above 90%) can be achieved. In addition, this study also found that while increasing the number of pathological images can improve the accuracy for new images, acquiring more patches from a single pathological image is ineffective in improving the accuracy for new images. This thesis provides directions for improving the pathological image classification accuracy while reducing the burden of pathologists on labeling data and reducing the computation time and cost of classifier training.