Enhanced Embolic Signal Analysis through Ensemble Deep Learning Techniques Utilizing Transfer Learning and Layer Freezing Strategies

Abstract

In the circulatory system, the presence of embolic particles, which are larger than the red blood cells, is one of the major causes of stroke. Hence, early and reliable detection of these particles is crucial in preventing potential adverse outcomes. Therefore, in this proposed study, 400 Doppler ultrasound signals that belong to three different classes (Speckle, Artifact, and Embolic) are examined for the detection of embolic signals (ESs). Each signal is transformed into a spectrogram image by using the short time Fourier transform, and the proposed learning models are fed by these images called spectrograms. The proposed architecture is developed as a fusion of 10 pretrained Convolutional neural network models, in which the transfer learning and freezing layer approaches are employed. In the fusion of models, the soft and hard voting methods are utilized as the ensemble learning approach. The obtained results show promising performance, achieving a classification accuracy of up to 96.73% and an F1 score of 96.5%. The findings of the study reveal that the proposed ensemble architecture has a high contribution in enhancing the detection of ESs, offering significant implications for stroke prevention strategies.