Photographic Analysis and Machine Learning for Diagnostic Prediction of Adenoid Hypertrophy

EasyChair Preprint 756

Abstract

Physiognomy has long been recorded in ancient Greece and ancient China. It predicts a person's character and health through facial features because some diseases’ traits may illustrate in face. Based on this, we apply a multidisciplinary method to investigate face appearance in photograph, identify adenoidal face, and early intervene in nasal respiratory obstruction. Using computer vision in feature selection, we identified most salient feature points of adenoid face including lip thickness, inner and outer eye distances. Through machine learning techniques, predictive models are constructed to discriminate adenoid face and non-adenoid face. The model-based analytical methods this article employed included decision tree, support vector machines, KNN and XGBoost. The reliability of forecasts was assessed by 5-fold cross validation. Two specific challenges were addressed in the study: Challenge 1, solve problem of head orientation and different illumination direction; Challenge 2, identify relevant facial prediction features which could convert into regression problem; Our research suggests that, compared to other approaches, computer vision feature selection provides a more reliable outcome forecasting of adenoids face, for example with a KNN cluster accuracy of about 77.41%.

Keyphrases: Adenoid Face, Data Mining, machine learning

@booklet{EasyChair:756,
  author    = {Xu Hu and Qinyan Zhang and Ji-Jiang Yang and Jiali Wu and Qing Wang and Yi Lei},
  title     = {Photographic Analysis and Machine Learning for Diagnostic Prediction of Adenoid Hypertrophy},
  howpublished = {EasyChair Preprint 756},
  year      = {EasyChair, 2019}}