Leveraging Global Mask for Structure Segmentation in Medical Images
W Lu1*, M Chen2, X Gu3, M Kazemimoghadam4, (1) UT Southwestern Medical Center, Dallas, TX, (2) UT Southwestern Medical Center, Dallas, TX, (3) Stanford University, Stanford, CA, (4) Dallas, TX
Presentations
PO-GePV-M-277 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Deep learning models for medical image segmentation are mainly intensity-based and thus highly influenced by intensity variations of input images, resulting in lack of generalization across modalities. Information regarding organs’ anatomical position and shape is relatively consistent across human subjects and could be learned and used for certain applications. We introduced a framework leveraging recurring anatomical patterns in the form of global binary masks to encode and utilize positional and shape information for medical image segmentation.
Methods: We studied two scenarios: (1) Global binary masks served as the only input for the segmentation model (i.e., U-Net). Hence, the network was forced to exclusively learn organs’ anatomical position and shape information for segmentation. (2) Global binary masks were incorporated as an additional channel functioning as position clues for the model. Two datasets of brain and heart structures were split into (26:10:10) and (12:3:5) for training, validation, and test respectively for model evaluation. The 2nd scenario was evaluated on small subsets of training data as well. The results were then compared to the baseline model trained on CT images only.
Results: In scenario 1, average dice scores of 0.82±0.03 and 0.77±0.06 were obtained across structures for the heart and brain datasets when global masks served as the only input. In scenario 2, for the model trained on two cases, utilizing the global masks as position clues improved the dice coefficients from 0 to 0.63±0.1 and 0.85±0.05 for segmenting the right eye in brain dataset and left atrium in the heart dataset respectively. Similar results were observed for different number of training data.
Conclusion: The findings imply that leveraging global binary masks is a promising direction for building segmentation models that are robust to image variations, generalizable across image modalities, and is an effective approach to compensate for scarcity of training data.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email