A Unified Deep Learning Model

Abstract

ABSTRACT newline newlineThe need for developing a Deep Learning model to extract robust features efficiently is a newlinevery important challenge as the extracted features can be used for classification and newlinesegmentation. This research work focuses on developing a novel Convolutional Neural newlineNetwork (MIDNet18) architecture which could handle different image types. Existing newlineresearch works clearly show that the traditional classification models are complex in newlinestructure, and hence their computational complexity is high. Moreover, these models work newlineless appropriately for sensitive medical datasets. To address the above said challenges, newlinethere is a need for a novel model (especially for medical dataset) with simple architecture newlinefor accurate classification and prediction. newlineMost research implementations were done on the publicly available datasets. Only a few newlineworks involved in the implementation of architecture for real time images. Along with newlinethose publicly available datasets, real time images were also used in this research work. newlineMoreover, there is a very small fraction of the research community involved in dental AI. newlineThe usage of a novel dental dataset containing the original maxillary photographs of the newlinepatients is also one of the novelties in this work. The annotation of all the collected dental newlineimages were done hand in hand with the dental experts in setting up the ground truth. newlineIn this research work, three different Convolutional Neural Network models namely newlineMIDNet11, MIDNet18-D200 and MIDNet18 were constructed from scratch and these newlinemodels were compared with each other for their performance on accurate classification. newlineThe MIDNet18 CNN model was shortlisted based on its excellent performance. This newlinemodel has been compared with various standard architectures (VGG16, VGG19, newlineResNet50, LeNet 5, AlexNet, DenseNet, and MobileNet) for different medical datasets. newlinenamely Lung, Retinal, and Brain datasets involving both binary and categorical newlineclassifications. newline