An efficient multi-stage ensemble deep learning framework for diagnosing infectious diseases

This study presents an efficient four-stage ensemble deep learning framework for diagnosing infectious diseases. The model is evaluated on three standard datasets. In our proposed four-stage transfer learning-based deep neural architecture (4s-min-FN), the images pass through four stages, each attempting to classify images as positive. A negative class is confirmed if every stage classifies the image as negative. This model (4S-min-FN) ensures the minimization of false negatives. When the new cases go through a changing scenario, the same model is modified (4S-min-FP) to minimize false positives following the same architecture but with a different transition rule. We use an adaptive threshold setting in the proposed architecture to find a proper trade-off between sensitivity, specificity, and good accuracy. We use well-known pre-trained deep neural architectures like Inception, ResNet-50, DenseNet-121, and MobileNet for the four-stage experimental evaluation and predicted the class, which provided better insights about the condition. The proposed model can perform at par with the existing techniques in terms of accuracy while reducing false positives and negatives depending on the requirement.