Multi-scale feature similarity-based weakly supervised lymphoma segmentation in PET/CT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of Diffuse Large B-Cell Lymphoma (DLBCL) in PET/CT images is crucial for prognosis.
• Manual DLBCL labeling is labor-intensive and time-consuming due to lesion variability.

Purpose of the Study:

• To develop a weakly supervised deep learning method for automatic lymphoma segmentation.
• To reduce the need for extensive, accurately labeled datasets in DLBCL imaging analysis.

Main Methods:

• A 3D V-Net with an Atrous Spatial Pyramid Pooling (ASPP) module was employed for multi-scale feature extraction.
• Weak supervision was achieved by using limited accurately labeled data alongside minimally labeled data.
• Cosine similarity was utilized to enforce multi-scale feature consistency between predicted and labeled regions.

Main Results:

• The proposed weakly supervised method achieved an average Dice Similarity Coefficient (DSC) of 71.47%.
• Performance was comparable to fully supervised methods when using a mix of accurately and weakly labeled data.
• The method demonstrated effectiveness in reducing the requirement for expert annotations.

Conclusions:

• Weakly supervised learning offers a viable approach for lymphoma segmentation, significantly reducing annotation burden.
• This method can improve the efficiency of DLBCL prognosis evaluation through automated segmentation.