Segmentation of medical devices
Segmentation of medical devices for minimally invasive surgery
Customer: Boston Children's Hospital (Boston, United States)
Summary: During minimally invasive cardiac surgery, it is critical for surgeons to have an accurate understanding of the location of medical devices such as catheters. However, a major challenge in localization is the limitations and image quality of the modality used during the procedure. Three-dimensional echocardiography, or ultrasound, is commonly used to visualize medical devices because it is a more flexible and cost-effective approach compared to magnetic resonance imaging and computed tomography. However, speckle noise and low resolution in ultrasound images can make it difficult to locate the object. In addition, catheters often have a curved path and can appear irregular on images, especially when the distal end of the catheter is perpendicular to the ultrasound probe. To address the problem of segmenting objects on images with speckle noise, we proposed a modification of the U-net architecture called V-net. The key feature of V-net is the additional skip connections within the encoder and decoder, which partially solve the problem of vanishing gradients and improve the segmentation accuracy as measured by the Dice Similarity Coefficient. The proposed V-Net was found to be 10% more accurate than a classical U-Net. The source dataset for this study was obtained using epicardial three-dimensional echocardiography during cardiac surgery on three Yorkshire pig hearts at Boston Children's Hospital. A catheter was inserted into the left ventricular cavity and the transthoracic X7-2t sensor was placed on the epicardium at the left ventricular apex. We used the Philips iE33 ultrasound machine and PMS5.1 ultrasound software to acquire 75 three-dimensional ultrasound gray-scale images of 176x176x208 voxels each. The catheter is poorly visible to the human eye on the echocardiographic data, so it is highlighted with green circles and ellipses in the samples shown.
Note: This project is related to the project "Segmentation of anatomical structures".
Project type: Commercial / Research
Media: PhD thesis, PhD thesis abstract
Figure 1. Catheter and pigtail used during the surgery
(a) Axial view
(b) Sagittal view
(c) Coronal view
Figure 2. Source data of patient 1
(a) Axial view
(b) Sagittal view
(c) Coronal view
Figure 3. Source data of patient 2
Figure 4. The architecture of the proposed fully 3D neural network
Figure 5. An approach used for feature transfer over encoder/decoder
(a) Axial view
(b) Sagittal view
(c) Coronal view
Figure 6. Segmentation of patient 1
Red mask is the ground truth segmentation, white mask - network prediction
(a) Axial view
(b) Sagittal view
(c) Coronal view
Figure 7. Segmentation of patient 1
Red mask is the ground truth segmentation, white mask - network prediction
(a) 3D segmentation
(b) 3D + time segmentation
Figure 8. Segmentation of 3D series
Red mask - network prediction
Figure 9. Comparison of the proposed V-net with existing solutions