Weakly supervised semantic segmentation for tuberculosis lung cavity diagnosis

Tuberculosis is a worldwide disease that threatens human health, and its early diagnosis is critical for effective treatment. The lung cavity is an important indicator for TB diagnosis, and its detection can provide valuable diagnostic information about tuberculosis lesions. However, traditional supervised learning methods for lung cavity detection usually require large amounts of labeled data, and obtaining these data is a time-consuming and laborious task for tuberculosis images. To address this challenge, a weakly supervised method for lung cavity semantic segmentation is proposed. In this approach, EfficientNet is utilized for co-training with image-level multi-class classification labels to generate regions of interest related to lung cavities. These generated regions are subsequently refined to determine the locations of lung cavities. This research results show that CT images under weak supervision method effectively segment lung cavity lesions. which achieves good performance without pixel-wise full supervision (W), with IoU and DSC of 31.2 % and 44.7%, respectively. It shows that weak supervision methods are in performance and even beyond some fully supervised learning methods.

Text 127292.pdf - Published Version Download (637kB) Official URL or Download Paper: https://www.ukm.my/apjitm/article/66b0c17acb2c7992...

Actions (login required)

View Item