三維深度卷積神經網路分類老鼠皮膚癌之光學同調斷層掃描影像

Abstract

本篇論文使用全域式光學同調斷層掃描 (Full-Field Optical Coherence Tomography, FF-OCT) 所獲得之三維微米級細胞組織影像，以監督式學習和三維卷積神經網路 (Convolutional Neural Network, CNN) 來實現小鼠鱗狀細胞癌 (Squamous Cell Carcinoma, SCC) 的影像分類，我們採用經典的神經網路作為三維模型架構的參考指標，並觀察模型的泛化能力和嘗試進行模型可視化。此外，我們更進一步評估和比較二維和三維卷積神經網路在光學同調斷層掃描影像中對小鼠皮膚病變的分類性能，探討透過三維卷積神經網路分析FF-OCT醫學影像的可行性，研究結果表明，雖然三維模型能在測試集上得到很好的分類性能，但礙於目前數據量的不足，在準確率的表現上較為不穩定。

The purpose of this study is to classify mice skin lesions in Full-Field Optical Coherence Tomography (FF-OCT) images using three-dimensional (3D) convolutional neural network (CNN). Deep CNN and supervised learning were implemented to extract features and achieve multi-class classification. We employed classic neural network architectures are trained with FF-OCT images as a reference indicator for 3D model designing. In addition, we evaluated and compared the performance of two-dimensional (2D) and 3D CNN applied to FF-OCT images to explore the feasibility to analyze FF-OCT images via 3D deep learning architecture. Research findings show that 3D CNN is effective for classification of mice squamous cell carcinoma (SCC) in test set, but the accuracy performance of 3D CNN is more unstable due to the influence of weight initialization and the amount of data.