Convolutional neural Network-based detection of deep vein thrombosis in a low limb with light reflection rheography

Abstract

Artificial intelligence has been widely used in the biomedical engineering field, which can assist the clinicians in disease diagnoses, help the engineers in physiological signal processing, or serve the people with chronic diseases in the homecare management. Blood clots in the deep veins of human body is called the deep vein thromboses (DVT). If the embolus passes through the lung, patient will have a life-threatening risk. Therefore, how to use the artificial intelligence to serve daily monitoring of the DVT condition is a valuable exploration. The light reflection rheography (LRR) has been used to detect the DVT of low limbs. In the previous study, the wearable device using LRR technique has been developed. But, this examination system could not be used by non-physician because the signal-quality evaluation of LRR and the classification of positive or negative DVT using the LRR signal all need the manual process. The goal of this study is to use a two-dimension convolutional neural network (2D CNN) to evaluate the qualities of LRR signals and classify the positive or negative DVT from the LRR signal with high reliability. The LRR signal and the smoothed signal were combined together to form a 450x450 image as the input pattern. In this study, twenty subjects were recruited to perform four-time experiments. A cuff pressured to 100 mmHg and 150 mmHg occluded the veins of low limbs to simulate the slight and serious DVT scenarios, and which was placed at the top and bottom of the knee of left leg to simulate the distal and proximal embolization. In the signal-quality evaluation, there were 700 samples including 476 high qualities and 224 low qualities, which were marked by the experts according to the vein emptying phenomenon. In the DVT classification, there were 476 samples including 167 negative samples, 158 slight positive samples, and 151 serious positive samples. A 19-layer CNN model proposed by Visual Geometry Group (VGG-19) was used in the two experiments. We performed the inter-group and intra-group analysis. Both results were better than the previous study. The accuracies of signal-quality evaluation and DVT classification were 0.92 and 0.75, respectively. Thus, the proposed method could support people with the high risk for DVT examination at non-medical settings.