Arefin Ittesafun Abian ¹ Mohaimenul Azam Khan ² Asif Karim ^2* Sami Azam ² Nur Mohammad Fahad ¹ Niusha Shafiabady ² Kheng Cher Yeo ² Friso De Boer ²

• ¹ United International University, Dhaka, Dhaka, Bangladesh
• ² Charles Darwin University, Darwin, Australia

An automated computerized approach can aid radiologists in the early diagnosis of lung disease from video modalities. This study focuses on the difficulties associated with identifying and categorizing respiratory diseases, including COVID-19, influenza, and pneumonia. We propose a novel method that combines three-dimensional (3D) models, model explainability (XAI), and a Decision Support System (DSS) that utilizes lung ultrasound (LUS) videos. The objective of the study is to improve the quality of video frames, boost the diversity of the dataset, maintain the sequence of frames, and create a hybrid 3D model (Three-Dimensional Time Distributed Convolutional Neural Network-Long short-term memory (TD-CNN-LSTM-LungNet)) for precise classification. The proposed methodology involves applying morphological opening and contour detection to improve frame quality, utilizing geometrical augmentation for dataset balance, introducing a graph-based approach for frame sequencing, and implementing a hybrid 3D model combining time-distributed CNN and LSTM networks utilizing vast ablation study. Model explainability is ensured through heatmap generation, region of interest segmentation, and Probability Density Function (PDF) graphs illustrating feature distribution. Our model TD-CNN-LSTM-LungNet attained a remarkable accuracy of 96.57% in classifying LUS videos into pneumonia, COVID-19, normal, and other lung disease classes, which is above compared to ten traditional transfer learning models experimented with in this study. The eleven-ablation case study reduced training costs and redundancy. K-fold cross-validation and accuracy-loss curves demonstrated model generalization. The DSS, incorporating Layer Class Activation Mapping (LayerCAM) and heatmaps, improved interpretability and reliability, and PDF graphs facilitated precise decision-making by identifying feature boundaries. The DSS facilitates clinical marker analysis, and the validation by using the proposed algorithms highlights its impact on a reliable diagnosis outcome. Our proposed methodology could assist radiologists in accurately detecting and comprehending the patterns of respiratory disorders.