Evaluation of Facial Paralysis Based on Spatial Features of Filtered Images

Abstract

Facial paralysis is a common clinical condition occurring in 30 to 40 patients per 100,000 people per year in Japan. A quantitative tool to support medical diagnostics is necessary. This paper presents a technique that we combined Gabor filters and wavelet decomposition to develop this tool. In our work, the Gabor filters and the wavelet decomposition are used as preprocessing steps to extract the feature. These features are used as the inputs of a multi-class support vector machines for quantitative evaluation of facial paralysis. Our method overcomes the drawbacks of the other techniques such as noisy removal and against variation of illumination. Experimental results show that our proposed method outperforms other conventional techniques testing on a dynamic facial expression image database.

A Review on Automated Facial Nerve Function Assessment From Visual Face Capture

“…This contrast manifests clear evidence of asymmetry, thus should not be excluded. To this end, analogous measures [30], [38], [44], [48]- [50] driven by pixel intensities as those used in static measures could be applied. A simple solution is to perform a subtraction between images obtained at rest and during facial movement, then compare the luminance changes of a specific paretic area with that of the healthy side [48], [49].…”

“…Such kind of methods however is sensitive to illumination changes, which is restricted to environment with controlled lighting. To cross this constraint, some studies resorted to robust visual texture descriptors [30], [38], [44], [50]. He et al [30] employed the multi-resolution LBP (MLBP) on temporal-spatial domain to extract the motion features from each region of the face.…”

“…Then they assessed the symmetry of facial motion by the Resistor-Average Distance (RAD, a distance measure between two probability distributions that is closely related to the Kullback-Leibler distance) between MLBP features. NGO et al further extended the facial texture analysis from spatial domain to frequency domain by using Gabor filters [50], circular Gabor filters [44]. More recent studies [14], [51] turned to deep learning methods such as convolutional neural networks (CNNs) which have revolutionized the visual imagery analysis to extract high-level features from the face image.…”

“…For a new subject, the solution first extracts computational features from his/her facial movement data, then calls a pre-trained classifier to map the features to the facial nerve function grade defined in the grading scale. The most frequently used grading scale is House-Brackmann scale (HBS) which divides the facial nerve function into six levels [11], [12], [30], [51], followed by Yanagihara scale (YGS) [44], [50] and Sunnybrook scale (SGS, see Table I) [47]. The grade could also simply be a binary value indicating whether the subject has facial palsy or not [37], [46], or if a specific face region is paralyzed [14].…”