Page 71 - Read Online

Page 162 Kimbowa et al. Art Int Surg 2024;4:149-69 https://dx.doi.org/10.20517/ais.2024.20

- Merge segmented patches to show the needle localization in
the 3D volume
- Centers of detected circular segmentations correspond to the
needle shaft
- Most distal bright intensity corresponds to the needle tip
[74]
2020- Lee et al. 2D - Pass image through model to get segmentation - Method evaluated on human data (8 - Evaluation metrics reported in # of pixels, rather than mm
01-20 - Apply a max contour algorithm to find most contiguous patients) - Only compared general segmentation architectures, but no
segment earlier needle detection methods
- Visualize by drawing bounding from top right most pixel to
bottom left pixel, and straightening the segmentation as
diagonal of the bounding box
[65]
2019- Mwikirize et al. 2D+t - Enhance needle tip in consecutive us images - Real time (67 fps) - Not robust to motion artifacts such as breathing, or
10-10 - Classifiy enhanced images and localize tip in enhanced images - Both in plane and out of plane detection pulsating
that have needle - Robust to intensity variations - Cannot detect stationary needle tip as it depends on motion
- Resilient to high intensity artifacts in the
image
- Incorporates temporal information
[90]
2019- Pourtaherian et al. 3D - Slice 3D ultrasound volume into 2D slices - Conceptually simple architecture - Computationally expensive
02-24 - Select 3 consecutive slices (with the middle one being the
reference slice) to incorporate some 3D information
- Pass slices as 3 channel input to fully connected network
(autoencoder style)
- Obtain pixelwise classification of the slices
[73]
2019- Arif et al. 3D+t - Segment 3D ultrasound volume using a CNN - Incorporates temporal information - Not robust to motion artifacts as it assumes only needle
02-11 - Extract needle candidates from segmentation using connected - Ablation studies performed on architecture moves between two consecutive frames
component labelling and PCA - Evaluated on multiple datasets (3 datasets) - Assumes linear needle motion
- Combine needle candidates with those from previous time - performance doesn’t vary much except for - Not robust to transducer motion (translation or rotation)
step to obtain real needle by detecting motion between the the in vivo data - Large standard deviation on in vivo data as compared to
time steps phantom data (not easily generalizable)
- Visualize needle in two planes; 1 perpendicular and the other - Computational speed measured on GPU
parallel to the transducer - Doesn’t localize needle tip - just the plane and segmentation
of the needle (perhaps visible shaft)
[72]
2018- Pourtaherian 3D - Extract voxels from 3D ultrasound volume and classify each - Can detect very short needles (5mm and - Method does not explicitly detect the needle tip (only the
05-31 voxel as needle or background 10mm) plane where the needle and tip are maximally visible)
- Obtain 2D cross section slices from the 3D ultrasound volume - Robust to transducer and patient - Can’t detect needle in the first 2mm
and segment the needle in each slice (various slices movements (as it performs repeated - Computationally expensive- Only in-plane
perpendicular to the lateral and elevation plane) detection in 3D volume)
- Map segmentation output onto its corresponding position in - Method evaluated on data from 2
3D transducers (of varying resolution) and 2
- Estimate needle axis by fitting a model of the needle to the tissue types, 2 needle types (of different
segmented voxels (model is a straight cylinder having a fixed gauge)
diameter)
- Visualize the 2D cross section plane that contains the entire
needle
[83]
2018- Mwikirize et al. 2D (1) Detect needle using a bounding box - Relies on intensity invariant features. - Inference time evaluated on GPU (most ultrasound devices
03-06 (2) Use bounding box to automatically determine needle Robust to low intensity needle features and run on CPU)
trajectory and tip presence of high intensity artifacts