Segmentation of anatomical structures
Segmentation of anatomical structures on 3D data
Customer: Boston Children's Hospital (Boston, United States)
Summary: We have proposed an ML-based solution for the automated quantification of several anatomical structures, including the left atrium, pancreas, spleen, hippocampus, and liver. Our solution uses a fully 3D neural network, known as a V-net, which incorporates additional skip connections to partially overcome the problem of vanishing gradients. We compared our solution to 11 existing approaches in the Medical Segmentation Decathlon competition and found that the V-net achieved the best segmentation accuracy (Dice Similarity Coefficient) for the pancreas (0.83), hippocampus (0.95), and liver (0.96). In addition, it had the second-best accuracy for the left atrium (0.92) and the seventh-best accuracy for the spleen (0.94). This project is related to the project "Segmentation of medical devices for minimally invasive surgery".
Collaborators: Nikolay Vasilyev (Pfizer, Boston, United States), Maria Ledesma-Carbayo (Universidad Politécnica de Madrid, Madrid, Spain)
Project type: Commercial / Research
Media: PhD thesis, PhD thesis abstract
Figure 1. The architecture of the proposed fully 3D neural network
Figure 2. An approach used for feature transfer over encoder/decoder
(a) Axial view
(b) Sagittal view
(c) Coronal view
Figure 3. Left atrium segmentation
Red mask - ground truth segmentation, White mask - network prediction
(a) Axial view
(b) Sagittal view
(c) Coronal view
Figure 4. Pancreas segmentation
Red mask - ground truth segmentation, White mask - network prediction
(a) Axial view
(b) Sagittal view
(c) Coronal view
Figure 5. Spleen segmentation
Red mask - ground truth segmentation, White mask - network prediction
(a) Axial view
(b) Sagittal view
(c) Coronal view
Figure 6. Hippocampus segmentation
Red mask - ground truth segmentation, White mask - network prediction
(a) Axial view
(b) Sagittal view
(c) Coronal view
Figure 7. Liver segmentation
Red mask - ground truth segmentation, White mask - network prediction