Page 64 - Read Online

Page 158 Kimbowa et al. Art Int Surg 2024;4:149-69 https://dx.doi.org/10.20517/ais.2024.20

[85]
efficiently detect the needle . Object detection can also be performed directly for the needle tip instead of
[86]
the entire needle .
DISCUSSION
Various methods and approaches have been developed to enhance needle alignment, visualization, and
localization in ultrasound. Currently, all methods that are aimed at improving needle alignment are
hardware-based [Figure 2], and require additional hardware which increases the cost of the ultrasound
systems and also disrupts normal workflow. The same applies to the hardware-based needle visualization
methods. This can be a big challenge, especially in resource-constrained communities such as rural and
remote settings, and low- and middle-income countries that can not afford additional hardware.

On the other hand, software-based methods do not require additional hardware and could be a potential
alternative in such scenarios. Classical image-based methods for needle visualization and localization rely
heavily on carefully engineered feature extractors and classifiers which are often not robust to various image
acquisition settings and image quality. Learning-based methods address this challenge by automatically
learning the feature extractor and/or classifier from existing data. Deep learning-based methods exhibit
superior performance compared to classical methods, and thus, this discussion will mainly focus on the
most recent deep learning-based methods for needle visualization and localization, summarized in Table 1
and detailed in Table 2.

The challenge with learning-based methods is that they require a lot of data to be trained. This data can be
collected from tissue-mimicking phantoms, freshly excised animal cadavers, or in vivo during clinical
procedures. Most methods use data collected by performing needle insertions in vitro with phantoms, and
ex vivo with porcine, bovine, and chicken while mimicking clinical scenarios [Table 1]. Only methods
developed for HDR prostate brachytherapy consistently use human in vivo data for evaluation. Future
methods can find motivation from Gillies et al., who evaluated their approach on in vivo datasets from
multiple organs and scenarios on top of the phantom datasets .
[89]
The data are typically annotated by an expert sonographer who performed the needle insertion experiments
to obtain the ground truth labels. In some scenarios, a hardware-based tracking system is used to obtain a
more accurate needle tip location, especially for cases where the needle is imperceptible to the human
eye [54,66,88] . In all the proposed approaches, local datasets were collected and this is not ideal for comparing
the proposed methods as noise and biases can easily be introduced into the data. To date, there is no
benchmark dataset on which developed methods can be evaluated, which has significantly stifled
progress .
[11]

The typical evaluation metric for learning-based methods is needle tip error, as the ultimate goal for needle
localization is to avoid puncture of critical tissue such as veins along the needle trajectory. For segmentation
methods that also detect the needle shaft, needle trajectory/orientation error is an important metric, on top
of the needle tip error, to assess model performance for needle guidance during insertion. Needle
localization performance has progressed over the years, in terms of needle tip localization error, up until
2022, when there seems to be a decrease in progress [Figure 4]. However, needle orientation error is also
used to ensure that a large portion of the needle shaft is also accurately detected. Out of all the proposed
methods, deep learning-based methods that report both tip localization and orientation error achieve state-
of-the-art performance [Figure 4B] [76,78,83] . Another key metric for software-based methods is inference time
on central processing unit (CPU), given that when deployed, these algorithms should achieve real-time
[14]
performance, which is considered to be any processing speed greater than 16 fps .