Background: Computer vision is a research field where algorithms are developed so that computers can gain high-level understanding from digital images or videos. With the rapid development and application of deep learning method, it has become one of the hottest area in the field of artificial intelligence. In clinical research, medical images are commonly used for disease diagnosis, which is a perfect application area for computer vision methods. In recent years, deep learning methods are widely applied in different types of medical images, which greatly improves the image-based disease diagnosis. In this thesis work, we focus on one of the dermatologic diseases, rosacea, to identify the subtypes based on facial images. With previous works utilizing deep neural networks on other medical image data achieving great performance, such as skin cancer and retinopathy screening, it is reasonable to apply deep neural networks on the facial images for rosacea disease prediction.
Methods and Materials: Facial images from rosacea patients with subtypes: ETR, PPR, PhR were collected as raw data. The raw image data were preprocessed to crop (solely facial region cropped from raw image) and mask (decoloring unnecessary region from cropped image) data. Each dataset were used under all the proposed models to evaluate the effect of image preprocessing. A simple 5-layer convolutional neural networks (CNN) was constructed as baseline model for disease prediction. Transfer learning from existing deep neural networks including ResNet, Inception, Inception-ResNet model were used to evaluate the prediction performance. To train and evaluate the model performance, 80% of each dataset were used as training set, 10% as validation set and 10% as testing set for final performance evaluation.
Results: Baseline CNN does not perform well on the current dataset with slightly higher than 50% of validation accuracy. Using transfer learning on all the deep neural network models has good performance on all three datasets, with worst performance occurs when using raw data, indicating the necessity of image preprocessing. ResNet152 and Inception-ResNet V2 were selected for disease prediction with highest validation accuracy of 93.6% and 93.95%, respectively, on the crop data. The final performance of ResNet152 and Inception-ResNet V2 on crop testing set had 85.84% and 93.24% testing accuracy, respectively.
Conclusion: Using transfer learning based on Inception-ResNet V2 model has achieved the best prediction performance on rosacea disease prediction with a 93.24% testing accuracy. Deep neural network architectures including ResNet or Inception can also be considered for dermatologic disease prediction with moderately good performance. The application of convolutional neural network on medical image analysis and disease diagnosis is promising and can be considered to extend to other medical area with image data analysis.
Table of Contents
About this Master's Thesis
|Subfield / Discipline|
|Committee Chair / Thesis Advisor|
|Dermatologic Diseases Prediction Using Deep Learning Method on Facial Images ()||2019-04-04||