胃食道逆流疾病分類之可詮釋性深度學習模型比較

Comparison of Interpretable Deep Learning Models for Classification of Gastroesophageal Reflux Diseases

邱愉晴(Yu-Ching Chiu)

指導教授 ： 曾明性

中山醫學大學/健康管理學院/醫學資訊學系/碩士(2021年)

https://doi.org/10.6834/csmu202100057

摘要

胃食道逆流疾病是常見的消化道疾病之一，以洛杉磯分類標準來評估內視鏡檢查嚴重程度。在長期胃食道逆流的影響之下，導致其他嚴重併發症的風險提高，例如巴瑞特氏食道、食道癌等。因此，應及早診斷出胃食道逆流的症狀，並予以適當的治療。
深度學習模型應用於電腦輔助診斷的案例越來越多，但是深度學習的「黑盒」特性，讓診斷結果難以解釋如此判斷的原因，妨礙醫師運用此方法進行病症的理解。在本研究中，使用上消化道內視鏡檢查影像作為資料集，可詮釋性模型為深度分層語義卷積神經網路，提供兩種輸出資訊，分別為語義特徵、以及洛杉磯分類標準預測，語義特徵為本研究人工萃取的3個內視鏡影像特徵，用以解釋模型分類預測結果的原因。接著比較語義特徵之網路層輸入融合層對於後續分類預測的準確性差異。泛化率最佳模型在十折交叉驗證平均下的訓練準確率為99.8%±0.2%，驗證準確率為85.8%±2.2%，測試準確率則是84.4%，高於兩位培訓醫師的人工測試結果（75%與65.6%）。

關鍵字

胃食道逆流疾病分類 ； 深度學習 ； 卷積神經網路 ； 可詮釋模型

並列摘要

Gastroesophageal reflux disease (GERD) is one of the common digestive tract diseases. The endoscopic severity evaluates by using the Los Angeles classification. Under the influence of long-term GERD, the risk of other serious complications increased. Such as Barrett's esophagus (BE) and esophageal cancer. Thus, diagnosed symptoms of GERD as early as possible and give appropriate treatment. More and more computer-aided diagnose using the deep learning model, but the deep learning "black box" feature makes it difficult to explain the reasons for such judgments for the diagnosis results. Endoscopists are hard to use this method to understand symptoms. In this study, a data set using Esophagogastroduodenoscopy images. And the interpretable model is a hierarchical semantic convolutional neural network (CNN). It can provide two kinds of output information, semantics features, and Los Angeles classification prediction. Semantic features are three endoscopic image features manually extracted in this study to explain the reasons for the model classification prediction results. Then compare the accuracy of the semantic features network layer input concatenate layer for after classification predictions. Under the 10-fold cross-validation, the training accuracy average of the best generalization model is 99.8%±0.2%, and the validation accuracy average is 85.8%±2.2%. The test accuracy is 84.4%, which is higher than the manual test results of the two trained physicians (75% and 65.6%).

並列關鍵字

gastroesophageal reflux disease classification ； deep learning ； convolutional neural network ； interpretable model

參考文獻

[1] Caruana, R., Lawrence, S., Giles, C. L. (2001). Overfitting in neural nets: Backpropagation, conjugate gradient, and early stopping. Paper presented at the Advances in neural information processing systems.

[2] Clevert, D.-A., Unterthiner, T., Hochreiter, S. (2015). Fast and accurate deep network learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289.

[3] Das, N., Topalovic, M., Janssens, W. (2018). Artificial intelligence in diagnosis of obstructive lung disease: current status and future potential. Curr Opin Pulm Med, 24(2), 117-123. doi:10.1097/MCP.0000000000000459

[4] Dent, J., El-Serag, H. B., Wallander, M. A., Johansson, S. (2005). Epidemiology of gastro-oesophageal reflux disease: a systematic review. Gut, 54(5), 710-717. doi:10.1136/gut.2004.051821

[5] Furneri, G., Klausnitzer, R., Haycock, L., Ihara, Z. (2019). Economic value of narrow-band imaging versus white light endoscopy for the diagnosis and surveillance of Barrett's esophagus: Cost-consequence model. PLoS One, 14(3), e0212916. doi:10.1371/journal.pone.0212916