Facial paralysis image analysis for stroke detection using deep ensemble transfer learning and optimization

Abstract

Facial paralysis (FP) weakens facial muscles, leading to asymmetric facial actions and complicating stroke diagnosis. Machine learning (ML) and deep learning (DL) systems have been explored for diagnosing FP, but the effectiveness of these methods is hindered by the limited size and diversity of available datasets. This study proposes a novel deep ensemble transfer learning method for accurate stroke diagnosis using facial paralysis imaging (DETLM-ASDFPI). The method leverages pre-trained models to reduce computation costs on edge devices. The framework includes data acquisition, preparation, and pre-processing, with image rescaling to standardize input dimensions. Feature extraction is performed using a deep capsule network (DCapsNet) to capture complex features. For stroke detection, an ensemble transfer learning model integrates three classifiers: gated recurrent unit (GRU), deep convolutional neural network (DCNN), and stacked sparse auto-encoder (SSAE). The hippopotamus optimization algorithm (HOA) is applied to fine-tune model parameters. The method was validated using two benchmark datasets, Massachusetts eye and ear infirmary (MEEI) and YouTube facial palsy (YFP), achieving an accuracy of 97.06%, outperforming recent approaches. This research demonstrates the effectiveness of the DETLM-ASDFPI method in accurately diagnosing strokes from FP images while addressing challenges related to dataset limitations and computational efficiency.