智慧化血液檢測系統之研究

Development of an intelligent blood test classification

黃宇業(Yu-Ye-Huang)

指導教授 ： 范憶華

中原大學/工學院/機械工程研究所/碩士(2019年)

https://doi.org/10.6840/cycu201900684

摘要

本文提出運用類神經網路來辨識血型的凝集程度，主要是應用正向分型測試及反向分型測試，將其運用在擁有6個微流道的光盤，同時直接在微流道中使用測定法，拍攝血液凝集照片進行影像處理。

辨識部分主要採用類神經網路來進行血液凝集判定，訓練資料為4000筆圖片；驗證集為400張，並且使用了深度學習的網路架構，來去作我們的結果判斷。首先將其作影像處理由RGB轉換為灰階，避免外在因素干擾判斷結果。神經網路內部做水平偵測、垂直偵測或是其他的特徵的卷積核，運算結束後會得出答案並運用Softmax函數為輸出，就可以得到血液凝集的分類型態還有該分類的信心值。

本研究執行20組病人血液(使用Epoch120數據來看)，成功辨識的為20組，成功率達100%，雖然深度學習神經網路會因為試片的關係沒辦法辨認出價數及血型，但對於血液的價數分類以及陽性及陰性判別毫無問題，也不會因此判別錯血型，大大保障了患者的安全。

使用Epoch120的數值相當穩定準確，算然速度較其他訓練慢，準確度比起Epoch20、40、60、80、100更好，辨識的成功率相對較高，還可能會勝過人類的肉眼視覺判別。

關鍵字

機器學習 ； 深度學習 ； 卷積神經網路 ； 影像處理 ； 血液凝集

並列摘要

Transferring blood, make sure that donors have a compatible blood type with the recipients. During the blood transfusion, if the donor’s blood is rejected by the recipient, a hemolytic transfusion reaction may occur. This may even be fatal.

Using neural networks to identify the degree of agglutination of blood types were presented in this paper. Forward typing and reverse typing tests were applied to the optical discs with 6 microchannels. Photos of agglutination in this paper were taken and processed.

For the identification part, the writer used deep learning to determine the degree of blood agglutination. The image was first converted to grayscale, so no external factors affected the result. The collected dataset was divided into three types: training set, verification set and testing set. After classifying the dataset, the training set was labeled and uploaded to the neural network for training, performing horizontal detection, vertical detection or other feature convolution kernel. After the operation, the answer was obtained by using the softmax function as output. Then, the type of the blood agglutination and the confidence value were shown on the user interface.

The writer examined 20 groups of patients with blood, In the study, 20 subjectsblood were examined. All of their blood were successfully identified, having 100 percent success rate. The deep learning neural network could not identify the number of bids and the blood type because of the optical discs; however, it could classify the number of bids, blood type and whether it’s positive or negative. This greatly ensure patients safety.

並列關鍵字

machine learning ； deep learning ； convolutional neural network ； image processing ； blood agglutination

參考文獻

[1] E. Kuriscak, P. Marsalek, J. Stroffek, and P. G. Toth, "Biological context of Hebb learning in artificial neural networks, a review," Neurocomputing, vol. 152, pp. 27-35, Mar 2015.

[2] R. Seising, "The Emergence of Fuzzy Sets in the Decade of the PerceptronLotfi A. Zadeh's and Frank Rosenblatt's Research Work on Pattern Classification," Mathematics, vol. 6, no. 7, Jul 2018, Art. no. 110.

[3] W. P. Jones and J. Hoskins, "BACK-PROPAGATION," Byte, vol. 12, no. 11, pp. 155-&, Oct 1987.

[4] A. Crisanti, D. J. Amit, and H. Gutfreund, "SATURATION LEVEL OF THE HOPFIELD MODEL FOR NEURAL NETWORK," Europhysics Letters, vol. 2, no. 4, pp. 337-341, Aug 1986.

[5] Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," (in English), Nature, Review vol. 521, no. 7553, pp. 436-444, May 2015.