以高低頻譜模組增進腸胃鏡檢測之效率

Abstract

自從機器學習因為硬體提升而開始蓬勃發展，尤其是深度學習，很多以往認為困難的任務，像是分類、影像切割、語意分析等，可以以更高的準確率及速度來完成。在醫療領域裡，長期需要大量人力去判斷病理，因此是非常適合機器學習發展的領域。感謝ACM Multi Media challenge 2019 釋出的腸胃鏡照片資料集，我們將實作深度學習模型去自動化分類16種腸胃圖片。 在台灣，腸道病例逐漸增加，且大腸癌歷年統計 [1] 的死亡率登上國人死因第三位，和醫生的討論亦得知食道炎、腸胃息肉的檢查為重要，需求也逐年增加，如何更快的發現病徵及提早治療為目前趨勢。在醫院裡，充滿著忙碌的醫生及病理研究員，檢查的任務很瑣碎卻又是必須做且要做得詳盡，例如細胞切片的判讀，必須從幾十億畫素的顯微鏡視野下找出些微的病理特徵，因此如果交由電腦判讀，即可快速掃整張玻片。而腸胃鏡檢測則以輔助判斷為目標，可提供經驗較少的醫生更好的意見，降低漏判發炎或息肉的狀況，增加可做腸胃鏡檢查的人手，AI著實有機會解決這問題。 此次研究中，以加速模型運算和視覺化，增加模型預測準確度為目標，並和醫生合作去尋找最好的模型。在深度學習領域中，有很多架構在大型資料集上有不錯表現，而我們將實驗Resnet152 [2]的架構，改用Octave convolution [3]並調整leaky relu的斜率，達成在CPU上2.71 FPS的速度及0.901的準確率。最後，期許這篇研究能促進深度學習在腸胃鏡追蹤的發展。

Since machine learning has been boosted significantly, especially for deep learning, by the development of hardware, a considerable amount of arduous task, like image classification and segmentation, can be solved by machine even with higher accurate and more robust than individuals. In the medical domain, there are a lot of works that need the more human source to complete, so it is suitable for applying ML. Thanks to ACM Multimedia challenge 2019 [4], we obtain an image dataset of “gastrointestinal track”(GI-track), which is the open-source of gastrointestinal track, and implement deep neural network to automatically classify the image. With the development of this task, I think the world can gain more benefits. In this thesis, we solve problems—improving model efficiency and robustness, and visualizing model interpretation—to predict label precisely. There are some solutions to solve the classification problem on large open datasets. However, we derived the ResNet152 model and changed its convolution layer with Octave convolution layer. To achieve training quickly, we apply leaky_relu with decay alpha to the activation function. As a consequence, the model can infer an image at speed 2.71 FPS on CPU and achieve 90.1% accuracy. Thus, we hope the development of machine learning in health care can be accelerated significantly.