Rare disease classification from facial photographs using deep learning
| dc.contributor.author | Ali, Hafsa Moontari | |
| dc.contributor.examiningcommittee | Leung, Carson (Computer Science) Turgeon, Max (Computer Science/Statistics) | en_US |
| dc.contributor.supervisor | Hu, Pingzhao (Biochemistry & Medical Genetics/Computer Science) Wang, Yang (Computer Science) | en_US |
| dc.date.accessioned | 2021-09-10T15:47:25Z | |
| dc.date.available | 2021-09-10T15:47:25Z | |
| dc.date.copyright | 2021-08-27 | |
| dc.date.issued | 2021-08-26 | en_US |
| dc.date.submitted | 2021-08-27T17:17:41Z | en_US |
| dc.degree.discipline | Computer Science | en_US |
| dc.degree.level | Master of Science (M.Sc.) | en_US |
| dc.description.abstract | Rare diseases affect a small number of populations all around the world. Often, rare disease patients are misdiagnosed and deprived of proper treatment due to the lack of knowledge about the diseases. The unavailability of standard data and methodologies to identify the rare diseases has made the situation more complex. Rare diseases are caused by malfunction of genes, and often leave noticeable traits on the face. In this thesis, a dataset of facial photographs of rare diseases is curated. The correlation between risk genes and facial features of rare diseases is calculated. Finally, rare diseases are classified from healthy facial photographs of children by employing transfer learning-guided pre-trained deep learning models. The performance of different convolutional neural network models (AlexNet, ResNet-18, ResNet-34, ResNet-50, VGG-16, VGG-19, DenseNet121, MobileNetV2) are analyzed over two different transfer learning approaches. All the models, except VGG-16, achieved superior results when trained with fine-tuned transfer learning approach than the other transfer learning approach where the convolution base was considered as a fixed feature extractor. The fine-tuned MobileNetV2 model showed the best classification result with 94.92% accuracy, and precision, recall, F1-score and AUC of 0.9498, 0.9492, 0.9475, 0.99, respectively. Then, augmentation is performed on ResNet-50 and DenseNet-121 and the overall performance improved in both transfer learning-based approaches. Two traditional machine learning based models (Support vector machine, eXtreme Gradient Boosting) are also applied for classification. The machine learning models achieved moderate results but underperformed comparing to the deep learning models. | en_US |
| dc.description.note | October 2021 | en_US |
| dc.identifier.uri | http://hdl.handle.net/1993/35960 | |
| dc.rights | open access | en_US |
| dc.subject | Rare diseases, Deep learning, Facial photographs, Gene similarity, Transfer learning | en_US |
| dc.title | Rare disease classification from facial photographs using deep learning | en_US |
| dc.type | master thesis | en_US |