基於 Inceptional 卷積神經網絡的醫學圖像分類 - DR眼底圖像

Medical Images Classification Based on Inceptional Convolutional Neural Networks - DR Fundus Images

洪千媛(Chien-Yuan Hung)

指導教授 ： 李昕潔

國立交通大學/管理學院/科技管理研究所/碩士(2020年)

摘要

隨者人工智能技術的快速發展，深度學習的應用已經成為跨學科的研究。深度卷積神經網絡(Deep Convolutional Neural Networks, DCNNs)更成為改善視覺辨識性能的重要工具，特別是在醫學圖像的應用上，更具有輔助醫生進行診斷的重要作用。近幾年，全球糖尿病人口已從4.25億逐年增加，因此，有必要將具有架構優化的進階卷積神經網絡應用到糖尿病視網膜病變(diabetic retinopathy, DR)圖像的分類任務。
當前糖尿病視網膜病變圖像的分類問題主要是使用傳統專業人工的手動特徵提取和遷移學習方法來進行分類預測。然而，普遍面臨過擬合、運算成本高和訓練時間長的問題並影響系統的有效運行。為了解決這些問題，本研究旨在通過適當的參數設置和潛在技能來優化原始的.SE-Inception.模型的架構，透過嵌入.Xavier-value.與.He-value.二種權重初始化方法，以及.PReLU.激活函數來保持神經元的活性並均勻分配權重。我們的實驗結果表明在有限的硬體設備與運算效能下，改良的SE-Inception.模型能減少參數量，提高收斂的速度，並避免過擬合發生的風險，達到較好且穩定的預測效果。

關鍵字

DR眼底圖像分類 ； 卷積神經網絡 ； Inception Net ； SENet ； PReLU ； Xavier-value ； He-value

並列摘要

While the progression of AI technology is on the increase, the application of deep learning has become an important issue as an interdisciplinary study. Deep Convolutional Neural Networks (DCNNs) have become a critical tool to improve the visual recognition performance, especially in medical images, which assist doctors in diagnosis. Moreover, the world's diabetic population in the past years has increased by 425 million. Thus, there is a need to apply CNNs with advanced structures on diabetic retinopathy (DR) image classification tasks. The DR image classification problem is mainly using traditional manual feature extraction and transfer learning methods. However, the problem of overfitting, high computing costs, and long training time are critical for the effective functioning of the system.
To address this problem, this study aims to optimize the structures of the original SE-Inception model with suitable parameter settings and latent skills by embedding the weight initialization of Xavier-value and He-value, and activation function of PReLU to maintain the neuron's activity and evenly assign weights. Our findings show that with limited hardware equipment and computing resources, the proposed improved SE-Inception model can reduce parameters, accelerate convergence, and avoid overfitting to achieve better and stable prediction results.

並列關鍵字

DR Fundus Image Classification ； Convolution Neural Networks ； Inception Net ； SENet ； PReLU ； Xavier-value ； He-value

參考文獻

[1] Patel, V. L., Shortliffe, E. H., Stefanelli, M., Szolovits, P., Berthold, M. R., Bellazzi, R., & Abu-Hanna, A. (2009). The coming of age of artificial intelligence in medicine. Artificial intelligence in medicine, 46(1), 5-17.

[2] Sahiner, B., Chan, H. P., Petrick, N., Wei, D., Helvie, M. A., Adler, D. D., & Goodsitt, M. M. (1996). Classification of mass and normal breast tissue: a convolution neural network classifier with spatial domain and texture images. IEEE transactions on Medical Imaging, 15(5), 598-610.

[3] Greenspan, H., Van Ginneken, B., & Summers, R. M. (2016). Guest editorial deep learning in medical imaging: Overview and future promise of an exciting new technique. IEEE Transactions on Medical Imaging, 35(5), 1153-1159.

[4] Guo, J., & Li, B. (2018). The application of medical artificial intelligence technology in rural areas of developing countries. Health equity, 2(1), 174-181.

[5] LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W., & Jackel, L. D. (1989). Backpropagation applied to handwritten zip code recognition. Neural computation, 1(4), 541-551.