利用移動窗口變換器級聯區域基底卷積神經網路於壓瘡偵測與診斷

Detection and Diagnosis for Pressure Injury by Using Swin-Cascade R-CNN

劉易鑫(Yi-Sin Liu)

指導教授 ： 張瑞峰

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202103625

摘要

壓瘡是一種發生率高的傷病，雖然致死機會不大但治療的代價高昂。為了給予壓瘡適當的治療，國際壓瘡諮詢委員會(National Pressure Ulcer Advisory Panel, NPUAP)將傷口依照侵入皮膚的嚴重度分從第一級至第四級。然而，不同傷口範圍及深度會影響判斷，甚至有出現診斷錯誤的可能性。因此需要一個電腦輔助診斷系統(Computer-aided Diagnosis System)協助醫護人員進行壓瘡的診斷。近年，卷積神經網路(CNN)被廣泛地用於診斷壓瘡。藉由專注在區域特徵，區域基底卷積神經網路(Region-based CNN, R-CNN)架構更是對偵測且診斷壓瘡具有優異的能力。由於區域基底卷積神積網路(R-CNN)架構的表現依賴於有效的特徵抓取，所以一個強力的特徵抓取模型為增強偵測及診斷壓瘡是需要的。
因此，為了設計一套電腦輔助診斷系統以進行壓瘡偵測及診斷，在這篇研究中我們使用了級聯區域基底卷積神經網路(Cascade R-CNN)與壓縮激勵移動窗口變換器(Squeeze-and-excitation Shifted Windows Transformer, SE-Swin Transformer)模型。此套電腦輔助診斷系統由資料擴增、特徵提取和傷口評級三部分組成。在資料擴增的部分我們調整了影像尺寸並增加影像，而後將經過擴增的影像送進我們提出的壓縮激勵移動窗口變換器(SE-Swin Transformer)做特徵提取。經過特徵擷取後，具階層式的特徵會被送進傷口評級產生出壓瘡偵測及診斷之結果。在傷口評級中，由於級聯區域基底卷積神經網路(Cascade R-CNN)可以產生一階階精準的偵測，因此將其作為我們的評級模型，最後產生壓瘡偵測與診斷之結果。此研究總共使用了883張壓瘡影像，其中包含240個第一級傷口，226個第二級傷口，203個第三級傷口和214個第四級傷口。根據實驗結果顯示，我們提出的系統能在偵測表現上達到平均精度均值81.3%，也能在診斷表現上達到準確率87.1%、靈敏度85.7%、陽性預測值86.6%的結果，顯示出我們提出的電腦輔助系統具良好的偵測與診斷能力。

關鍵字

壓瘡 ； 電腦輔助診斷 ； 注意力機制 ； 變換器 ； 區域基底卷基神經網路

並列摘要

Pressure injury was a chronic disease with high incidence. Although the fatality rate was low, the cost of treatment was high. To make appropriate treatments for the pressure injury, the National Pressure Ulcer Advisory Panel (NPUAP) ranked the injury from grade I to grade IV according to the severity of skin invasion. However, different injury ranges and depths would affect the evaluation of the pressure injury and make the incorrect diagnosis. Therefore, a computer-aided diagnosis (CADx) system was required to assist medical personnel in diagnosing pressure injury. In recent, the convolutional neural network (CNN) was widely employed in diagnosing pressure injury. By concentrating on the region of features, the region-based convolutional neural network (R-CNN) architectures even presented excellent ability to detect and diagnose pressure injury. Since the performance of the R-CNN architectures relied on effective feature extraction, a robust feature extraction model was required in improving the ability to detect and diagnose pressure injury.
Thus, to design a CADx system in detecting and diagnosing pressure injury, the Cascade R-CNN with the squeeze-and-excitation shifted windows transformer (SE-Swin transformer) model was employed in this study. The proposed CADx system consisted of data augmentation, feature extraction, and injury grading. In the data augmentation, the images were resized and augmented and then were sent to the SE-Swin transformer model for feature extraction. Through the feature extraction, the hierarchical features were extracted and fed into the injury grading, producing the detection and diagnosis results of the pressure injury. In the injury grading, the Cascade R-CNN was employed as a grading model since the three-stage shared heads could generate precise detection stage by stage. Ultimately, the detection and diagnosis results of the pressure injury were produced. In experiments, a total of 883 images, including 240 grade I injuries, 226 grade II injuries, 203 grade III injuries, and 214 grade IV injuries, were used in this study to evaluate the performance of our proposed system. The results showed that our proposed system could achieve 81.3% mAP in the detection performance, and it could obtain 87.1% accuracy, 85.7% sensitivity, and 86.6% PPV in the diagnosis performance. It indicated our proposed CADx system with well ability to detect and diagnose pressure injury.

並列關鍵字

Pressure injury ； computer-aided diagnosis ； attention mechanism ； transformer ； region-based convolutional neural network

參考文獻

[1] L. E. Edsberg, D. Langemo, M. M. Baharestani, M. E. Posthauer, and M. Goldberg, "Unavoidable pressure injury: state of the science and consensus outcomes," Journal of Wound Ostomy Continence Nursing, vol. 41, no. 4, pp. 313-334, 2014.

[2] C. VanGilder, C. Lachenbruch, C. Algrim-Boyle, and S. Meyer, "The international pressure ulcer prevalence™ survey: 2006-2015," Journal of Wound, Ostomy and Continence Nursing, vol. 44, no. 1, pp. 20-28, 2017.

[3] G. C. Xakellis, R. Frantz, and A. Lewis, "Cost of pressure ulcer prevention in long‐term care," Journal of the American Geriatrics Society, vol. 43, no. 5, pp. 496-501, 1995.

[4] D. Z. Bliss, B. L. Westra, K. Savik, and Y. Hou, "Effectiveness of wound, ostomy and continence–certified nurses on individual patient outcomes in home health care," Journal of Wound Ostomy Continence Nursing, vol. 40, no. 2, pp. 135-142, 2013.

[5] T. Zhao et al., "A hybrid CNN feature model for pulmonary nodule differentiation task," in Imaging for Patient-Customized Simulations and Systems for Point-of-Care Ultrasound: Springer, 2017, pp. 19-26.