Non-Subjective Assessment of Facial Anomalies

Every year an estimated 400,000 children are born with rare craniofacial anomalies that have a major effect on their quality of life. Early diagnosis and intervention will greatly increase these children's quality of life. The quantitative evaluation of facial features is difficult and subjective grading scales have well-known limitations. This study investigates the use of computer vision to help detect craniofacial abnormalities from non-anomalous facial images automatically. This thesis develops a non-subjective approach to efficiently and accurately characterizing the facial features of craniofacial anomaly patients and uses this technology to provide a quantitative and non-subjective measurement of anomaly as part of an automated computerized diagnostics device. Portrait images of non-anomalous faces with neutral expressions were taken from multiple databases containing 130 males and 80 females. Two face space models based on eigenfaces were developed, one for each gender category, for automatic facial feature localization in each face space. An anomaly image was projected onto the face space, and the intensity of the anomaly was determined by the rate of fairly projection. The accuracy of the algorithm was compared to manual ratings, and output of the model was quantified using non-anomalous face projection limits. The findings reveal that the proposed models are remarkably effective in assessing quantification of facial anomalies with only fewer number of images. Since only pixel values are considered throughout the processing; the personal identification, and descriptions of the input face images are preserved in this approach. However, the main challenges in this study are due in part to patient privacy issues and a lack of training data.