Improved Multi-modal Patch Based Lymphoma Segmentation with Negative Sample Augmentation and Label Guidance on PET/CT Scans.

Lymphoma is a cancer of the lymphatic system, and it can affect many organs throughout the body. Positron emission tomography (PET)/computed tomography (CT) are primary imaging methods to assess lymphoma types and monitor their treatment, where PET is sensitive to identify lymphoma regions while CT preserves anatomical structures. Combining PET and CT is thus useful for lymphoma segmentation because it helps to identify lymphoma types and evaluate treatment effects. However, lymphoma segmentation suffers many challenges, including substantial lymphoma size and shape variance, numerous types, limited PET/CT data for lymphoma, and similar PET signals with adjacent organs. To address these challenges, we integrate label guidance, patch sampling, and negative data augmentation to achieve multi-modal lymphoma segmentation. The training data consist of positive and negative patch samples. These samples are purposely extracted from the original scans with the guidance of lymphoma labels. Negative samples are further supplemented from the PET/CT scans of non-lymphoma patients to better discriminate lymphoma from adjacent organs. The proposed method was validated on the PET/CT scans from 28 patients. Experimental results revealed that the Dice coefficient was improved from 0.11 to 0.43 in comparison with a baseline method the 3D-residual U-Net method. Patch-based strategy is also computational undemanding. These results suggest that the proposed method could be an efficient means to segment lymphoma and possibly used for identifying lymphoma types and assessing their treatment.