Classification of Facial Images to Assist in the Diagnosis of Autism Spectrum Disorder: A Study on the Effect of Face Detection and Landmark Identification Algorithms

• Gabriel C. Michelassi USP
• Henrique S. Bortoletti USP
• Tuany D. Pinheiro USP
• Thiago Nobayashi USP
• Fabio R. D. de Barros USP
• Rafael L. Testa USP
• Andréia F. Silva USP
• Mirian C. Revers USP
• Joana Portolese USP
• Helio Pedrini UNICAMP
• Helena Brentani USP
• Fatima L. S. Nunes USP
• Ariane Machado-Lima USP

Resumo

Since facial morphology can be linked to brain developmental problems, studies have been conducted to develop computational systems to assist in the diagnosis of some neurodevelopmental disorders based on facial images. The first steps usually include face detection and landmark identification. Although there are several libraries that implement different algorithms for these tasks, to the best of our knowledge no study has discussed the effect of choosing these ready-to-use implementations on the performance of the final classifier. This paper compares four libraries for facial detection and landmark identification in the context of classification of facial images for computer-aided diagnosis of Autism Spectrum Disorder, where the classifiers achieved 0.92, the highest F1-score. The results indicate that the choice of which facial detection and landmark identification algorithms to use do in fact affect the final classifier performance. It appears that the causes are related to not only the quality of face and landmark identification, but also to the success rate of face detection. This last issue is particularly important when the initial training sample size is modest, which is usually the case in terms of classification of some syndromes or neurodevelopmental disorders based on facial images.