Pulmonary tuberculosis (PTB) is a contagious disease that affects the lung region. PTB is a life-threatening disease if it is detected late or left untreated. To perform the initial screening of PTB, the World Health Organization has recommended chest radiograph. Till now, the screening process requires either the patients to come to secondary health centers from rural areas or the radiologists to go the remote locations. This process is rejuvenated with the introduction of computer-aided diagnosis (CAD) systems. CAD systems reduce the need for expert radiologists in the screening process. However, the development and deployment are still in the early phases as new methods are being developed to improve the performance of CAD systems in terms of accuracy, specificity and sensitivity. In this study, a deep learning-based PTB classification system has been presented that achieves the state-of-the-art performance for TB classification. Firstly, a proposed architecture based on the blocks is presented and then it is used to create an ensemble. In the proposed ensemble, two standard architectures namely AlexNet, and ResNet have also been used in addition to the proposed architecture. All the architectures are trained and evaluated on a combined dataset formed using publicly available standard datasets. The proposed ensemble attains the accuracy of 90.00% and area under the curve equal to 0.96, which is better than the performance of the existing methods.
相似文献In this era of Internet, the exchange of data between the users and service providers has grown tremendously. Organizations in health, banking, social network, criminal and government sectors have been collecting and processing the individuals’ information for their gainful purpose. However, collecting and sharing of the individuals’ information which could be sensitive and confidential, for data mining may cause a breach in data privacy. In many applications, selective data collection of confidential and sensitive information of the users’ needs to be modified for preserving it from unauthorized access and disclosure. Many data mining techniques that include statistical, k-anonymity, cryptographic, perturbation and randomization methods, etc. have been evolved for protecting and preserving data privacy. These techniques have their own limitations, it may be the case that the privacy protection is adequate or computations complexities are high and expensive. To address the limitations of the above-mentioned techniques, a methodology comprising of encoding and randomization, is proposed to preserve privacy. This technique called as Randomized Encoding (RE) technique, in which encoding is performed with addition of random noise from a known distribution to the original data for perturbing the data before its release to the public domain. The core component of this technique is a novel primitive of using Randomized Encoding (RE) which is quite similar to the spirit of other cryptographic algorithms. The reconstruction of an approximation to the original data distribution is done from the perturbed data and used for data mining purposes. There is always a trade-off between information loss and privacy preservation. To achieve balance between privacy and data utility, the dataset attributes are first classified into sensitive and quasi-identifiers. The pre-classified confidential and sensitive data attributes are perturbed using Base 64 encoding with addition of a randomly generated noise for preserving privacy. In this variable dynamic proposed approach, the result analysis of the experiment conducted suggests that the proposed technique performs computationally efficient and preserves privacy while adequately maintaining data utility in comparison with other privacy preserving techniques such as anonymization approach.
相似文献