Dropout Regularization for Automatic Segmented Dental Images

Abstract

Deep neural networks are those networks that have a large number of parameters, thus the core of deep learning systems. There is a challenge that arises, from these systems as to how they perform against training data, and/or validation datasets. Due to the number of parameters involved the networks tend to consume a lot of time and this brings about a condition referred as over-fitting. This approach proposes the introduction of a dropout layer between the input and the first hidden layer in a model. This is quite specific and different from the traditional dropout used in other fields which introduce the dropout in each and every hidden layer of the network model to deal with over-fitting. Our approach involves a pre-processing step that deals with data augmentation to take care of the limited number of dental images and erosion morphology to remove noise from the images. Additionally, segmentation is done to extract edge-based features using the canny edge detection method. Further, the neural network used employs the sequential model from Keras, and this is for combining iterations from the edge segmentation step into one model. Parallel evaluations to the model are carried out, first without dropout, and the other with a dropout input layer of size 0.3. The introduction of dropout in the model as a weight regularization technique, improved the accuracy of evaluation results, 89.0% for the model without dropout, to 91.3% for model with dropout, for both precision and recall values.