Automatic lung tumor segmentation and classification based on improved U-Net assisted hybrid deep learning approach

Abstract

Early lung tumor detection is crucial because symptoms often appear late. The cloud revolutionize early diagnosis through real-time data collection and on-device model training. Machine learning and image processing techniques have already shown promise in identifying lung tumor from scans. Traditional methods struggled with less accuracy and less speed seems to be the major limitations. Early detection saves lives by enabling treatment before the cancer seals airways and infections set in. For the early diagnosis of lung tumor, the proposed methodology is developed with novel segmentation, feature extraction and classification models under improved U-Net model, stacked auto encoder and dilated depth-wise separable convolutional network assisted bidirectional long short term memory model (DDWC-Bi-LSTM) respectively. To segment the medical images, U-Net is improved with guided attention mechanism. To further validate the efficiency of proposed approach, the dataset is utilized in terms of accuracy, precision, recall, F1-score, and error. Proposed technique obtain an accuracy of 96.62% and 97.86% for kaggle and Medical Segmentation Decathlon lung datasets respectively.