Quantitative Detection Of Immunochromatographic Strips Based On Machine Learning

Immunochromatography is a new immunological detection technology that combines the specific immune response between antigens and antibodies with thin-film chromatography.As an important part of immunochromatographic technology,gold immunochromatographic strip(GICS)is widely used in medical diagnosis,food detection,environmental sanitation and other fields due to its advantages such as convenience,celerity,strong specificity and high sensitivity.In the current application,GICS is mainly used for qualitative or semi-quantitative detection because of the impact of the detection environment and its technical limitations.However,this detection method has certain limitations.Not only does it consume huge manpower,but it also provides large subjective errors as well as less detection information,causing it fails to meet the application needs.Its further promotion and use are limited to a certain extent.Consequently,there is a wide demand for quantitative analysis of GICS,which has become a future development trend.We aim to achieve accurate identification and quantitative detection of gold immunochromatographic strips through image processing.The main research contents are as follows:(1)Aiming at the characteristics and application requirements of GICS,the paper proposes to use support vector machine optimized by particle swarm optimization to segment and identify GICS,and to transform the problem of strip recognition into pixel classification problem.First,effective input features are constructed based on the information of the gray level and position of the strips.Second,considering that the penalty coefficient and nuclear control parameters in the support vector machine(SVM)have a great impact on model performance,we use the particle swarm optimization(PSO)algorithm to perform parameter adjustment to improve the model's classification performance.Finally,the function fitting relationship between the strip concentration and the relative integrated optical density value(RIOD)was established to achieve an quantitative analysis of the GICS.(2)In order to improve the recognition accuracy of GICS,the paper proposes a new multi-mode self-learning particle swarm algorithm(MSPSO).MSPSO can collect the particles' experienced information as a guide for the current evolution direction,and determine the mode of the particles at different stages.It can also select the optimal evolution strategy to achieve the balance between the local and global optimization capabilities of the particles,thus,improve the performance of the algorithm.We further constructs improved MSPSO algorithm to optimize the SVM model to realize the recognition of the target area of the GICS.The experimental results show that the improved MSPSO-SVM method can effectively improve the recognition accuracy and quantitative detection accuracy of the GICS.(3)To further improve the recognition efficiency and accuracy of the GICS image,we propose to extract the edge lines of the target area of the strip by using improved reinforcement learning method and creatively transform the strip recognition problem into pixels location strategy optimization problem.Additionally,aiming at the dimensional disaster problem in traditional reinforcement learning,we use deep belief network(DBN)to extract the state characteristics of strips and adopts multi-factor learning curve strategy to dynamically change the capacity of the replay memory and the sampling size to improve the efficiency of the algorithm.The results manifest that the improved deep reinforcement learning model proposed in this paper has higher efficiency and accuracy in segment and recognition of GICS,and has better application value.