Lung nodule classification using optimized sensitive deep convolutional neural network

Abstract

According to the findings of the National Cancer Registry Program Report 2020, released by the Indian Council of Medical Research (ICMR) in collaboration with the National Center for Disease Informatics and Research (NCDIR) Bengaluru, there will be approximately 13.9 lakhs malignant growth cases at the end of 2020, and that number will rise up to 15.7 lakhs by 2025. As a result, an in-depth, highly sophisticated system for the recognition and classification of lung nodules is required to reduce the mortality rate. Deep learning, an area of artificial intelligence, is one of the strategies for diagnosing cancer nodules using computer-aided diagnostic software. Deep neural network architectures have many layers that enables them to conceptualize features effectively on their own. Because of the enormous dimensionality of information, it is difficult to tune the boundaries of human expertise specifically for the lung nodule classification. Due to the lack of positive class sample availability, handling data imbalance can be a big challenge. For any dataset, hyperparameter tuning plays a significant role in determining the optimized model construction. So, for handling the automated feature extraction, data imbalanced problem and hyperparameter optimization of the network, different techniques have been proposed. newlineThe first technique employs Deep Convolutional Neural Network architecture for automated feature extraction and classification of lung nodule candidates as benign or malignant using cost-sensitive function and data augmentation techniques. The progressive scaling method was used in the second approach. The model was first trained with small images (n x n), then gradually increased the size of the images in the ratio of 2n x 2n, and then trained again until the maximum size of the image was not reached. In terms of enhancing classifier results, this approach appears to be promising. In the third method, efficient pruning and searching techniques were used to suggest an automated hyperparameter optimization t