大多數的醫學影像的數據集都會面對數據不平衡問題,主要因素是比起正向個案(異常個案),負向個案(正常個案)比較容易收集。更何況,醫學影像數據集都是屬於小型數據,而這些問題都是醫學影像數據集在深度學習領域的分類任務的最大挑戰。另一方面,生成式對抗網絡 (Generative Adversarial Networks, GANs) 已證明其生成的合成數據能在各種領域上使用,如修復影像、超解析度成像以及生成更高質量的影像。生成式對抗網絡所生成的合成數據可以增加原本數據集的數量,成為解決小型數據問題的方案。再加上,基於ImageNet所預訓練的模型是這個實驗模型的基礎訓練模型以得到更好的表現性能。但是,ImageNet數據集的領域與醫學影像的領域相差甚遠,因此透過使用生成式對抗網絡所生成的合成數據能幫助原本分類任務。這個實驗主要有兩大貢獻:提出一個有效提升性能表現的訓練系統、基於Grad-CAM++上進行量的測定以探討有關使用合成數據的影響。為了分析生成式對抗網絡所生成的合成數據對於分類任務的性能,在此使用可視化分析工具——Grad-CAM++並採用量的测定,即交並比 (Intersection over Union, IoU) 以分析深度學習模型在透過分類任務裡學到了什麼。此實驗也比對不同的訓練系統進行分析以從中發現模型如何學習和表現的關鍵。最後,透過Grad-CAM++的交並比結果、測量的準確率 (accuracy) 、精確率 (precision) 、召回率 (recall) 、綜合評價指標 (F1 score)以及曲面下的接收者操作特征 (Area Under the Receiver Operating Characteristics, AUROC)的結果,新的訓練系統——第四方法,會被提出如何合理使用合成數據來提升原本的分類性能。第四方法的訓練系統透過量的測定以驗證其實用性和表現性能,結果上是比其他傳統的訓練系統,在分類性能的整體上進步了許多。在分類性能的測量結果中,進步高達2.0228%;在交並比測量的結果中,進步高達8.2815%。
Most of the medical imaging datasets usually have to deal with imbalance problem, which show that negative case (normal case) is easier to collect and find compare to positive case (abnormal case). This issue and the small-scale medical imaging dataset are the biggest problems in training the deep learning classification task. In the other hand, Generative Adversarial Networks (GANs) has shown that it can generate synthetic data that is useful in various applications, such as reconstructing images, super resolution imaging, generating high quality of images. Synthetic data generated by GANs might be a solution to small-scale issue by increasing the size of dataset. Models pre-trained on ImageNet are used as base models to get better performance. However. the domain of ImageNet dataset is different from the domain of medical imaging. Therefore, the synthetic data generated by GANs based on medical imaging domain can be helpful. This experiment has two main contributions: proposing a effective training pipeline and analyzing the impact of synthetic data through quantity measurement on Grad-CAM++. To analyze the impact of GANs-based synthetic data on the downstream classification performance, quantity measurement: intersection over union (IoU) metric is applied on visual explanation tool: Grad-CAM++, to analyze what deep learning model has learned through the classification task. Various training pipelines are compared to gain insight into how the model learns and performs. From the observation results of both IoU metric measurement in Grad-CAM++ and measurement of accuracy, precision, recall, f1-score, and Area Under the Receiver Operating Characteristics (AUROC), a new training pipeline: method 4 is proposed to handle the synthetic data to improve the overall classification performance. The overall classification performance of the proposed pipeline is improved compare to the other training pipelines. The maximum improvement rate in classification performance is 2.0228%, where the maximum improvement rate in IoU metric is 8.2815%.