Developing lightweight cnn models for detecting lesions in the medical imaging

Abstract

Classification and characterization of the lesions in medical imaging are getting interested in the field of computer-aided diagnostics (CAD) systems for radiology. A num ber of machine learning (ML) algorithms are proposed for classification and characterization. All the existing models suffer from inter-observer variability. CNN models are based on Deep Neural Networks, which are distinct from typical Machine Learning methods such as k-NN, Decision-Trees, and so on. Moreover, the performance of Deep Neural Network-based approaches is better as compared to traditional Machine Learning approaches as these models extract the features from training data automatically. In the past, the researchers have proposed many CNN architectures such as VGG16, VGG19, Inception V3, MobileNet, ResNet50, etc. for the classification of 1000 class ImageNet datasets. These models can also be utilized for the classification of other datasets by transfer learning. While pre-trained CNN models work well, the high number of parameters required makes them computationally intensive. In this thesis, we proposed several simplified and optimized CNN models for the classification and characterization of the lesions in the carotid artery, lung, and brain (Wilson disease). The datasets are taken from the medical practitioners at different geological locations and several publicly available datasets. We got the ground truth from the radiologist for real-time data. Then we augmented the data into folds. In these cohorts, we ran all of the recommended models, including CNN models, Transfer Learning (TL) models, and Hybrid CNN models. We achieved an accuracy of 95.66% in carotid using optimized CNN and 99.45% accuracy using hybrid CNN, 98.9% in lung, and 91.2% in Wilson disease. We validated our models and hypothesis on the characterization of the multicenter study. We compared the performance of the AI models on various performance metrices and then ranked the models based on the grading schema for the best classifier. We tested the performance of the models on various cross-validation protocols and on different hardware resources. We achieved a prediction time on the local computer of less than 2 sec for all the datasets.