Özet:

Abstract The morbidity and mortality due to pneumonia is drastically increased in the developing countries due to multiple factors like poverty, poor health care, overcrowding, poor hygiene, malnutrition, and air pollution. It is more important to implement the effective diagnostic model, which can assist in detecting & differentiating pneumonia from other lung diseases. Usually, lung disease symptoms do not show up until the disease become severe. Hence, we propose a model that detects pneumonia early with the help of x rays and text data (symptoms and signs). Dataset & proposed model/methods: This work puts up an effort to model a new lung disease classification with multi-modal data, where the information of symptoms is in text form along with X-ray images. Initially, the dataset is collected manually with eight chest diseases in text and image format. And, the size of the dataset is 2286 x 3600 images. And, the steps like pre-processing, feature extraction takes place separately for both text input and image input. Further, improved feature level fusion is performed that combines both the features to determine the final classification. Particularly, statistical features, IG, and improved entropy features will be extracted from the text input. GLCM, MBP, and improved CSLBP features are extracted from the input image. Further, the fused features are subjected to the hybrid classification model that integrates the Deep Maxout and CNN to categorize the lung diseases. Optimal training is carried out in the hybrid classification model via a new CBRACDC (Customized Battle Royale Algorithm with Canberra Distance Calculation) algorithm.  Finally, the superiority of the proposed work is evaluated over the conventional models. Results: The proposed model attains an enhanced accuracy as 85.13%, 87.99%, 92.28% and 95.28 % for various learning percentages like 60, 70, 80 and 90 respectively than conventional models. It also achieves better sensitivity, specificity, f value as well as low FPR, FNR and error rate when compared to conventional models. The best result achieved at learning percentage 90.  Conclusion: It is concluded that the recommended model has an effective performance than the existing models.    

Anahtar Kelimeler: