RAPID-Net: Reduced architecture for pneumonia in infants detection using deep convolutional framework using chest radiograph

Abstract

ObjectivePneumonia is a thoracic disease triggered by a pathogen infection in the lungs. A routine diagnostic test to detect pneumonia is a chest radiography. Radiologists or expert physicians help to localize and diagnose pneumonia clouds in the chest X-Ray. Radiological practice is error-inclined due to the increasing patient volumes, limited number of experts and the subjective nature of individual perception. The existing computer aided pneumonia diagnostic techniques faces issues like imbalanced pneumonia dataset, limited quality of chest radiograph, pathological deformities, X-Ray imaging inhomogeneities, gross background noise, overlapped patterns of opacities and anatomical alterations caused due to misaligned body positioning. To mitigate the number of diagnostic flaws and to ease the task of radiologists, there is a necessity for computer-aided pneumonia detection from a given chest radiograph. MethodsReduced Architecture for Pneumonia in infants Detection (RAPID-Net) is introduced using thin stacking of convolutional layers having faster convergence along with suitable preprocessing and augmentation techniques addressing above mentioned challenges while extracting significant local and global statistical features for pneumonia cloud detection. ResultsThe retrospective quantitative analysis of the RAPID-Net in pneumonia detection reviews a promising accuracy rate of 94.6%, sensitivity score of 94.62%, and F1-score as 93.89%. SignificanceThe proposed RAPID-Net model highlights reduced deep architecture having 96% lesser trainable parameters and 42% reduction in computational time as compared to other competitive best performing pretrained backbone deep network. The clinical conduction and validation of results is supported by Radiologist, Sun Lab Diagnostic, Virar and Cardinal Gracias Memorial Hospital, Vasai.