Multi-task learning and weakly supervised semantic segmentation for classification and segmentation of pulmonary tuberculosis lesions in CT images

Tuberculosis (TB) remains a significant global health challenge, characterized by high incidence and mortality rates on a global scale. With the rapid advancement of computer-aided diagnosis (CAD) tools in recent years, CAD has assumed an increasingly crucial role in supporting TB diagnosis. Nonetheless, developing such tools for TB diagnosis heavily relies on well- annotated computerized tomography (CT) datasets. However, currently, the available annotations in TB CT datasets are still limited. For instance, the types of annotated lesions are relatively singular, which to some extent constrains the further development of CAD tools for TB diagnosis. To address this limitation, this research introduces DeepPulmoTB, a CT multi-task learning dataset explicitly designed for pulmonary TB comprehensive radiological feature analysis. This dataset contains three key diagnostic indicators for TB, which are: i) lung cavity, ii) lung areas, and iii) C-LCW (consolidation or LC wall). To demonstrate the advantages of DeepPulmoTB, this research proposes a novel comprehensive radiological feature analysis framework for TB imaging based on multi-task learning. This framework includes a multi-task learning model, namely DPTBNet, for the joint segmentation and classification of lesion tissues in CT images. The architecture of DPTBNet comprises two subnets, which are: i) SwinUnetR for the segmentation task, and ii) a lightweight multi-scale network for the classification task. In addition, to address the problem that traditional deep learning methods based on supervised learning paradigms need to rely on a large amount of accurate pixel-level data for the segmentation of lung cavity, this research constructs a weakly supervised semantic segmentation framework for lung cavity segmentation. Through this framework, radiologists need only to classify the number and the approximate location (e.g., left lung, right lung, or both) of the lung cavity in the CT scan to achieve efficient segmentation of this kind of lesion. This process eliminates the need for meticulously drawing boundaries, greatly reducing the cost of annotation. In this framework, to effectively improve the ability to identify lung cavity lesions, this research designs a novel CT medical imaging analysis model for TB, named SwinUNeLCsT. SwinUNeLCsT is a hybrid model comprising a convolutional neural network and transformer. This model aims to improve the ability to obtain lung cavity imaging features by efficiently integrating the local texture information extracted by the convolutional network and the high-level semantic information captured by the transformer. This research validates the effectiveness of the DPTBNet model and the practical value of the DeepPulmoTB dataset in TB diagnosis tasks through extensive experiments. In fully supervised segmentation tasks, SwinUNeLCsT outperforms current mainstream 3D medical segmentation models, achieving a 95% Hausdorff Distance (95HD) of 23.54 ± 6.86, Intersection over Union (IoU) of 0.365 ± 0.113, and Dice Similarity Coefficient (DSC) of 0.526 ± 0.121. Additionally, in the weakly supervised learning framework, SwinUNeLCsT demonstrates superior performance, achieving a 95HD of 31.43 ± 7.46, IoU of 0.267 ± 0.126, and DSC of 0.418 ± 0.134, outperforming existing 3D medical weakly supervised learning methods. These experimental results fully demonstrate the excellent applicability and developmental potential of the methodologies proposed in this study in the field of intelligent analysis of TB imaging.

Text FSKTM 2024 24 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (5MB) Text FSKTM 2024 24.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (7MB)

Actions (login required)

View Item