Brachial Plexus Nerve Trunk Recognition From Ultrasound Images: A Comparative Study of Deep Learning Models

Abstract

Brachial plexus block is a common regional anesthesia method widely used in upper limb surgery. Nowadays, ultrasound-guided brachial plexus block has been extensively used in clinical anesthesia. However, accurate brachial plexus block is highly dependent on the physician's experience, and a physician without extensive clinical experience may cause nerve injury when performing a nerve block. With the development of artificial intelligence technology, the deep learning method can automatically identify the brachial plexus in ultrasound images and assist doctors in completing the brachial plexus block accurately and quickly. In this paper, we aim to evaluate the performance of different deep learning models in identifying brachial plexus (i.e., segmentation of brachial plexus) from ultrasonic images to explore the best models and training strategies for this task. To this end, we use a new dataset containing 340 brachial plexus ultrasound images annotated by three experienced clinicians. Among the 12 deep learning models we evaluated, U-Net achieves the best segmentation accuracy, with an intersection over union (IoU) of 68.50%. However, the number of U-Net parameters is very large, and it can only process 15 images per second. Compared to U-Net, LinkNet can process 142 images per second and achieve the second-best segmentation accuracy with an IoU of 66.27%. It achieves the balance between segmentation accuracy and processing efficiency, which has a good potential for the brachial plexus's real-time segmentation task.

Artificial intelligence for ultrasound scanning in regional anaesthesia: a scoping review of the evidence from multiple disciplines

“…Early approaches, most published before 2016–17, typically involved steps of feature extraction from the ultrasound image, followed by classification of these features, for both PNB 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 and CNB. 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 Around this point, there has been an increase in published data where the dominant technique was deep learning for both PNB 9 , 11 , 12 , 38 , 39 , 40 , 41 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 , 50 , 51 , 52 , 53 , 54 , 55 , 56 , 57 , 58 , 59 , 60 , 61 , 62 , 63 , 64 , 65 , 66 , 67 , 68 , 69 , 70 , 71 , 72 , 73 and CNB, 7 , 74 , 75 , …”

“…In the available sources, a variety of human image interpretation was used as the ground truth. This included a range in the number of humans used in the ground truth from one, 7 , 23 , 24 , 25 , 31 , 32 , 34 , 35 , 36 , 38 , 46 , 49 , 58 , 60 , 67 , 72 , 77 two, 8 , 9 , 19 , 21 , 26 , 30 , 33 , 44 , 61 three, 51 , 52 , 53 , 63 , 65 and up to 15. 48 Some sources cited ‘multiple’ humans, used a plural term (e.g.…”

“…Sources described their ground truth human(s) as a volunteer, 64 doctor/physician, 48 , 52 , 63 , 65 sonographer, 7 , 30 , 31 , 32 , 33 , 34 , 35 , 60 , 75 , 77 clinical professional, 27 , 38 or (regional) anaesthetist. 8 , 17 , 18 , 19 , 20 , 21 , 23 , 25 , 26 , 36 , 44 , 45 , 46 , 49 , 51 , 53 , 54 , 58 , 61 , 62 , 67 , 69 The level of expertise of those providing the ground truth was described as trained, 64 experienced, 21 , 30 , 31 , 32 , 33 , 34 , 46 , 48 , 52 , 65 expert, 7 , 8 , 17 , 18 , 19 , 20 , 25 , 26 , 27 , 33 , …”

“…The most common output for models processing PNB ultrasound images was segmentation (the outline/colour overlay of features in the image). 9 , 11 , 12 , 19 , 20 , 22 , 23 , 25 , 27 , 36 , 38 , 39 , 40 , 41 , 42 , 43 , 45 , 46 , 47 , 48 , 49 , 50 , 51 , 52 , 53 , 54 , 55 , 57 , 58 , 59 , 60 , 63 , 64 , 65 , 66 , 67 , 68 , 69 , 71 , 72 , 78 , 79 Less common outputs included the use of bounding boxes to identify the location of the structure, 8 , 17 , 18 , 26 , 61 attempting to locate the centre of a nerve and placing expanding circular rings from the centre, 21 or placing name tags over structures in question. 9 , 43 , 45 , 47 , 50 …”

Optimal view detection for ultrasound-guided supraclavicular block using convolutional neural networks

“…2 However, successful SCB requires skilled ultrasound manipulation such as tilting the probe to obtain an optimal ultrasound view. 3 Previous studies showed the AI-based CADx systems for the segmentation and detection of the brachial plexus, or its identification at the interscalene level [4][5][6][7][8] , however, it simply focused on recognizing the location of anatomical structures on ultrasound. The proposed methods describe the development of deep learning-based CADx systems that can distinguish between optimal and non-optimal views for ultrasound-guided SCB.…”