基於深度學習分析的醫學影像中的結直腸息肉分類

Colorectal Polyps Classification in Medical Image Using Deep Learning Analysis

南帝奈(Natinai Jinsakul)

指導教授 ： 蔡正發

國立屏東科技大學/國際學院/熱帶農業暨國際合作系/博士(2021年)

https://doi.org/10.6346/NPUST202100104

摘要

結直腸癌（CRC）是一種對數百萬人來說越來越重要的醫學問題，並開始頻繁地出現在年輕人中。透過篩查方法預防大腸癌可以通過使用顯示人體各個部位的醫學圖像（例如結腸造影中的拓撲圖圖像）來檢測初期階段，以識別息肉的異常生長。這些圖像可以讓醫生知道在更嚴重的癌症發展之前切除息肉的位置。然而，在診斷過程中仍有許多局限 性。幸運的是，有一個解決辦法，即是通過使用醫學成像和電腦輔助診斷（CAD）和人工智慧（AI）的計算技術以進行分析。
本研究探討了通過人工智慧對深度學習的改進，在圖像分類的背景下訓練幾種卷積神經網路（CNN）演算法，使之適用於結腸造影過程中的智慧診斷系統，其主要目的是：首先，利用深度學習技術，在大腸定位圖像資料集預處理中生成圖像增強方法以進行分類；其次，比較原始圖像資料集和圖像增強資料集的訓練CNN性能；再者，採用Swish激勵函數代替傳統的激勵函數，以提高CNN的深度學習性能；之後，利用多種CNN技術訓練結直腸地形圖圖像資料集進行圖像分類；最後，比較多種CNN技術的分類性能，以尋找一個合適的模型來開發CAD系統在結腸造影中的應用。
結果顯示，大多數CNN模型在使用圖像增強資料集的結直腸位置圖像進行訓練時，可以比用原始數位資料集訓練時可以提高分類性能。不同的CNN模型用Swish激勵代替了傳統的激勵函數，結果顯示，在應用Swish激勵函數進行二類別和三類別分類時，並不是每個CNN都能提高分類性能。研究結果顯示Xception是CNN體系結構中最有潛力的圖像分類能力，通過測試結果可以獲得97.97% 的準確率，正確預測率為98.52%。在Swish激勵函數的輔助下，Xception的分類結果顯示，進行兩類別分類準確率提高到98.99%。通過正確測試，分類正確率達到99.63%，而進行三種類別的分類能力獲得了91.48%的準確率，而在測試階段，總分類的正確率為80.95%。

關鍵字

深度學習 ； 卷積神經網絡 ； 圖像預處理 ； 結直腸癌 ； 大腸息肉 ； 圖像分類 ； 地形圖圖像

並列摘要

Colorectal Cancer (CRC) is a cancer which is becoming an increasingly significant medical issue to millions of people and has begun to frequently appear in younger people. Prevention of CRC by a screening approach can work by detecting the initial stage by using medical images which demonstrate various parts of the human body, such as topogram images in colonography to identify abnormal growth of polyps. The images allow physicians to know the coordinates to remove the polyps before a more severe form of cancer develops. Nevertheless, there are numerous limitations in the diagnosis procedure such as manual interpretation being tedious, the lengthy time consumption, and the potential for bias and individual error. Fortunately, there is a solution through employing medical imaging with an analysis of computational technology based on Computer-Aided Diagnosis (CAD) with Artificial Intelligence (AI).
This study investigates the improvement via AI of deep learning by training several Convolutional Neural Network (CNN) algorithms in the context of image classification to be suitable for intelligent diagnosis systems in the colonography process in the main purposes: firstly, to generate an image augmentation method in pre-processing the colorectal topogram image dataset for classification with deep learning techniques; secondly, to compare the training CNN performance between the original image dataset and the image augmentation dataset; thirdly, to apply the Swish activation function that can improve the deep learning of CNN performance rather than the traditional activation function; fourthly, to employ several CNN techniques in training with the colorectal topogram image dataset for image classification, and finally, to compare the classification performance of several CNN techniques and discover a suitable model for developing the application of a CAD system in the colonography procedure.
The main concept of this study inspired by several CNN architecture modification parts for enhancing the image classification performance, which focused on the optimization of image augmentation method the colorectal topogram image data set and applying the new activation function to replace the traditional activation function. When all processes were finished, the experimental results were compared to find the suitable model. The results revealed that most CNN models when training with the colorectal topogram images of the image augmentation dataset can improve the performance better than when training with the original number dataset for classification into two classes. The different CNN models, by replacing the traditional activation function with Swish activation, show results which compare to show that not every CNN can improve classification performance when applying the Swish activation function for classifying into two classes and three classes. The results indicated that Xception is the CNN architecture that generated the most promising image classification capability for training with the image augmentation dataset by obtaining accuracy of up to 97.97% by the testing results, representing correct prediction of 98.52%. Xception, with the assistance of the Swish activation function, produced results which illustrated that the classification performance for classification into two classes increases the classification accuracy to 98.99%. By testing correctly, classification achieved 99.63%, while classification into three classes indicated that the classification ability gained accuracy of 91.48%, while for the testing stage it was established that the total correctly classifies 80.95%.

並列關鍵字

deep learning ； convolutional neural network ； image preprocessing ； colorectal cancer ； colorectal polyp ； image classification ； topogram image

參考文獻

Abadi, M., A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, and M. Devin. 2015. Tensor-Flow: Large-scale Machine Learning on Heterogeneous Systems. arXiv arXiv:1603.04467.

Alom, M. Z., M. Hasan, C. Yakopcic, T. M. Taha, and V. K. Asari. 2020. Improved Inception-Residual Convolutional Neural Network for Object Recognition. Neural Computing and Applications 32(1):279-293.

American Cancer Society (ACS). 2020. Colorectal Cancer Facts & Figures 2020-2022. Retrieved on Jul. 23, 2020, from the World Wide Web: https://www.cancer.org/content/dam/cancer-org/research/cancer-facts-and-statistics/colorectal-cancer-facts-and-figures/colorectal-cancer-facts-and-figures-2020-2022.pdf

An, F., and Z. Liu. 2019. Medical Image Segmentation Algorithm Based on Feedback Mechanism Convolutional Neural Network. Biomedical Signal Processing and Control 53:101589.

Anwar, S. M., M. Q. Majid, A. Adnan, A. Muhammad, K. Majdi, and K. Muhammad. 2018. Medical Image Analysis Using Convolutional Neural Networks: A Review. Journal of Medical Systems 42(11):226.