機器學習輔助影像分割用於提升晶片上細胞檢測通量

Machine Learning-Based Image Segmentation for Enhancing the Throughput of Cell Evaluation on a Chip

張仁碩(Jen-Shuo Chang)

指導教授 ： 蔡佳宏

國立交通大學/工學院/機械工程系所/碩士(2020年)

摘要

本研究提出了一種自動紅血球細胞影像分割的方法，用於識別高密度紅血球細胞流中的每個細胞影像，進而提高微流體晶片中對紅血球機械性質檢測的通量。一般晶片上細胞檢測常用影像處理去做自動的細胞辨識，因為在高密度細胞之間影像互相重疊，導致無法分割出每個細胞影像並追蹤其動態，所以只能透過稀釋細胞密度來進行檢測間接地限制了檢測通量。本研究提案的分割方法包含了標記的分水嶺演算法以及特徵點切割法，自動分割高密度的紅血球細胞影像，再以機器學習驗證分割結果，達到高通量的細胞檢測。實驗中首先準備了4種不同濃度的稀釋血液，全血含量分別為2%、4%、10%和20%，其中2%血液濃度為過去在晶片上檢測常用的濃度。透過提案方法可以使10%濃度追蹤率從24%提高至96.4%，相較於傳統使用2%血液濃度的檢測通量提高6倍; 此外執行了不同輸入壓力的實驗，分別為101.7kpa、102kpa、102.2kpa和102.7kpa，對於不同壓力影響不同流速的實驗結果，說明此壓力範圍內的流速不影響本實驗對細胞機械性質的檢測。本研究利用機器學習方法開發晶片上高密度紅血球追蹤系統，可提高晶片上紅血球細胞檢測通量，預期能助益於細胞大數據資料蒐集，提供細胞研究與臨床醫學使用。

關鍵字

機器學習 ； 影像分割 ； 晶片 ； 細胞檢測通量

並列摘要

This study proposes an automatic red blood cell (RBC) image segmentation method to identify single RBCs in a high-density RBC stream, thereby enhancing the throughput of cell evaluation on a chip. Generally, image processing is employed for on-chip cell evaluation. It is difficult to separate single RBCs from high-density RBC stream because the RBC images are often overlapped. Therefore, evaluation can only be performed by greatly diluted blood samples, which limited the throughput of the evaluation. In this study, a novel segmentation method, which includes the labeled watershed algorithm, the feature point segmentation, and machine learning classifier SVM, is proposed for automatically identifying single RBCs from high-density RBC images. The method is expected to achieve high-throughput RBC evaluation. In the experiment, 4 different concentrations of blood are tested, and the concentrations are 2%, 4%, 10%, and 20%, respectively. The mechanical properties of RBC in different concentrations are evaluated. The proposed method increases the tracking rate with the blood concentration of 10% from 24% to 96.4%, and it enhanced the evaluation throughput by 6 times respect to conventional evaluation with 2% blood sample. In addition, 4 different driven pressures, which are 101.7kpa, 102kpa, 102.2kpa, and 102.7kpa are tested for different flow rates in experiments. The results show that the flow rate does not affect the evaluation of cell mechanical properties in this range of pressure. To sum up, a machine learning based segmentation method is proposed and is applied in automatic image tracking for high-density RBC on a chip. The method could benefit in the big data of cell properties as well as assisting diagnosis in medical applications.

並列關鍵字

Machine Learning ； Image Segmentation ； Chip ； Throughput of Cell Evaluation

參考文獻

[1] J.Stuart andG. B.Nash, “Red cell deformability and haematological disorders,” Blood Rev., vol. 4, no. 3, pp. 141–147, 1990.

[2] H. L.Reid, J. A.Dormandy, A. J.Barnes, P. J.Lock, andT. L.Dormandy, “Impaired Red Cell Deformability in Peripheral Vascular Disease,” Lancet, vol. 307, no. 7961, pp. 666–668, 1976.

[3] A.Alizadeh andA.Dadvand, “Effects of deformability of RBCs on their dynamics and blood flow passing through a stenosed microvessel: an immersed boundary-lattice Boltzmann approach,” Theor. Comput. Fluid Dyn., vol. 32, no. 1, pp. 91–107, 2018.

[4] “Mokken, Kedaria, Henny - 1992 - Annals of Hematology.pdf.” .

[5] A. M.Dondorp, P. A.Kager, J.Vreeken, andN. J.White, “Abnormal blood flow and red blood cell deformability in severe malaria,” Parasitol. Today, vol. 16, no. 6, pp. 228–232, 2000.