Improving Early Detection of Lung Disorders: A Multi-head Self-Attention CNN-BiLSTM Model: Fid-Bau Portal

Improving Early Detection of Lung Disorders: A Multi-head Self-Attention CNN-BiLSTM Model

Indumathi, V. / Siva, R.

Respiratory system diseases are a leading cause of increased mortality, morbidity, and disability rates globally. Lung disorders occur due to constant exposure of the lungs to harmful agents present in the ambient air. Early diagnosis is the only prevention measure to avoid the risk of health issues. In recent times, numerous technologies exist to treat pulmonary disease in the early stages. However, evaluating pulmonary microstructural variations is considered a challenging task in routine clinical practice. In this paper, we propose a novel automated pulmonary disease prediction system using the convolutional neural network-based bidirectional neural network (CNN-BiLSTM) with multi-head self-attention layer approach to address the above-mentioned issue. We implemented this novel approach on four medical image datasets, namely, chest X-ray 14, contrast-enhanced CT (CECT), PET/CT, and the Japanese Society of Radiological Technology (JSRT) databases. The proposed approach predicts diseased instances from the image datasets and classifies them as normal cases and predicted diseases of multiple classes, which include pneumonia, lung cancer, COVID-19, asthma, etc. To attain high classification accuracy, we preprocess the datasets using normalization and standardization pipelines. The existing methods such as DLSA, VDSNet, SLIC, and DNN models offer accuracy of 92.5%, 93.7%, 93.2%, 93.5%, and 91.8%, 93.8%, 94%, and 91% on the chest X-ray and JSRT datasets, respectively. The proposed CNN-BiLSTM with multi-head self-attention layer approach offers an accuracy of 94.9%, 97.8%, 97.9%, and 94.6% in the chest X-ray, PET/CT, CECT, and JSRT datasets, respectively.