Myocardial Amyloidosis Detection With Terahertz Spectroscopy And Neural Networks

The combination of terahertz time-domain spectroscopy technology and deep learning methods has been widely used in the field of biomedicine,providing new ideas for the intelligent detection of myocardial amyloidosis.However,the implementation of this technology still exists some challenges: the irregularity of tissues,the spectral overlap in different kinds of tissues,and the achievement of classifying a small number of spectra with low signal-to-noise ratio.On this basis,this paper first introduces the background of myocardial amyloidosis detection using terahertz time domain spectroscopy and artificial intelligence.Then,it elaborates on the efficient construction of spectral datasets and further describes the structure and performance analysis of this classification model.Finally,to be closer to the actual application scenarios,this paper elaborates the lesion detection process in the scenario of low signal-to-noise ratio and small samples.Specifically,the main contents and innovations of this paper are as follows:1.An edge spectral rejection algorithm for biological samples is proposed to realize the efficient construction of terahertz standard spectral datasets.Myocardial tissue is extremely precious,so the original information of the samples should be preserved as much as possible.However,due to the small size and irregular shape of the tissue,it is inevitable that mixed spectra of "non-sample region and sample area" will be collected,which will affect the purity of spectral information and interfere with subsequent classification of spectra.Through the analysis of spectral trends,a data cleaning algorithm based on the first-order differential threshold is designed to realize the automatic rejection of edge spectra,and a standard dataset with a number of 4319 is constructed.2.A sensitive and real-time deep learning model was designed for myocardial amyloidosis detection.The acquired dataset which includes four different signal-tonoise ratios is taken from tissue with three thicknesses,so how to weaken the difference of spectra within classes and enlarge the differentiation of spectra between different classes is particularly important.According to the characteristics of the dataset,we design a convolutional neural network with multi-scale feature extraction module and dense connection module.An accuracy of 98.37%,a precision of 97.94%,a recall of98.85% and a F1-score of 98.39% are achieved.The time required to classify a spectrum is only 0.0035 seconds,and the time consumption of overall detection process is within1 hour,which has exceeded the efficiency of existing clinical methods.3.We realize the detection of myocardial amyloidosis in scenarios with low signal-to-noise ratio and small samples,which greatly compresses the data acquisition time.According to the dataset containing 100 spectra averaged for 1 time,we propose a deep learning model which is composed of a convolutional noise reduction autoencoder,a multi-scale feature extraction module and a dense connection module.Besides,an accuracy of 95.00%,a precision of 100%,a recall of 92.30% and a F1-score of 95.99% are achieved.The time to obtain and classify a spectrum is within 1second and 0.004 seconds,respectively.Thus,the overall time consumption is controlled at about 100 seconds,which reduces the cost of time greatly.In addition,by combining the proposed model with the data amplification algorithm,the detection of myocardial amyloidosis under the condition of "unbalanced,low signal-to-noise ratio and small samples" is achieved.