Detection of Pulmonary Disorders with CNN model using Knowledge Distillation technique

Chronic respiratory conditions are a major public health concern worldwide and pose a significant burden on individuals and healthcare systems. Early detection of pulmonary diseases play a major role in treatment. However, existing diagnostic techniques can be expensive, time-consuming, and require specialized expertise and any misdiagnosis can lead to delayed treatment. Lung auscultation is a widely used diagnostic technique that involves listening to respiratory sounds through a stethoscope which provide valuable information about the respiratory system and help in the identification of pulmonary diseases. However, interpretation of these sounds requires extensive training and experience. In this paper, we present the ”Pulmonet” system, which has been developed to identify pulmonary diseases by analyzing lung auscultation sounds. The methodology employed involves knowledge distillation, where valuable insights from a large and computationally intensive model (referred to as the teacher model) are transferred to a computationally efficient model (referred to as the student model). Our approach involves converting the respiratory sounds into Mel spectrogram images, which can be processed by our system. The model’s task is to make predictions about the presence of pulmonary diseases in a patient and identify the specific disease if any (such as Pneumonia, Upper Respiratory Tract Infection, Chronic Obstructive Pulmonary Disease, Bronchitis, Bronchiolitis) or classify them as healthy if no disease is detected. Following that, the insights acquired by the initial model are conveyed to the student model, which possesses lower computational complexity, making it feasible for deployment on mobile devices. Finding the suitable loss function for student model is crucial and challenging in our work. The loss function should be designed to balance both knowledge transfer and preservation of student model. In our work we have achieved the accuracy of 82% in teacher model and 78% in student model.