Investigation of Appropriate Scaling of Networks and Images for Convolutional Neural Network-Based Nerve Detection in Ultrasound-Guided Nerve Blocks

Takaaki Sugino, Shinya Onogi, Rieko Oishi, Chie Hanayama, Satoki Inoue, Shinjiro Ishida, Yuhang Yao, Nobuhiro Ogasawara, Masahiro Murakawa, Yoshikazu Nakajima2024 · DOI

TL;DR

This paper tackles the challenge of accurately identifying nerves in ultrasound images for regional anesthesia, a crucial but difficult task due to image noise and variability. The researchers propose using a Scaled-YOLOv4 deep learning model and investigate how different model sizes and input image resolutions impact detection performance, finding that larger models with smaller input images yield the best results for both accuracy and real-time processing.

What Problem Does This Paper Address?

Ultrasound imaging is indispensable for guiding medical procedures like peripheral nerve blocks (PNBs), which are used to numb specific areas of the body. However, visualizing nerves in ultrasound images is tricky. Factors like speckle noise, acoustic shadows, and the inherent variability in nerve appearance make it hard for even experienced anesthesiologists to pinpoint nerves accurately. This difficulty can lead to suboptimal anesthesia delivery or even nerve damage. While deep learning, particularly Convolutional Neural Networks (CNNs), has shown promise in image analysis, there's a need to optimize these models for the specific challenges of ultrasound nerve detection, especially concerning how network complexity and input image characteristics affect performance and real-time usability.

Methods

The core of this study lies in investigating the optimal configuration of a deep learning model for nerve detection. The researchers adopted Scaled-YOLOv4, a state-of-the-art object detection framework known for balancing speed and accuracy. They specifically explored different model scaling factors (varying the number of convolutional layers and filters) and input image scaling (resizing images to different resolutions).

The Scaled-YOLOv4 architecture consists of a backbone for feature extraction, a neck for feature refinement, and a head for predicting bounding boxes. The study utilized several variants of Scaled-YOLOv4 (CSP, P5, P6, P7) to represent different levels of model complexity. For training, they employed standard techniques like pre-trained weights on ImageNet, a combination of CIoU loss for bounding box regression and cross-entropy loss for classification, and data augmentation (translation, scaling, flipping, mix-up). Predictions were made using a confidence score threshold and non-maximum suppression.

Datasets

The study utilized two datasets:

  • Public Dataset: A publicly available dataset from Kaggle containing 619 ultrasound images of the sciatic, ulnar, femoral, and median nerves. These images were acquired with a resolution of 640x480 pixels and cropped to a region of interest of 360x279 pixels. Each image contained only one nerve.
  • Original Dataset: A newly compiled dataset comprising 993 ultrasound images of the brachial plexus from 101 healthy volunteers. These images were captured with a higher resolution (1024x768 pixels) using a different device and included not only nerves but also surrounding structures like blood vessels and muscles. This dataset aimed to provide richer contextual information for discrimination. Images were cropped to various region of interest sizes (540x753, 600x800, 605x710 pixels).

Key Insights and Ideas

  • Model Size Matters, But So Does Input Resolution: The study highlights a critical interplay between network complexity and input image size. While larger, more complex models (like YOLOv4-P7) generally improve accuracy by capturing more intricate features, the resolution of the input image has a surprisingly strong counteracting effect.
  • Smaller Images Can Be Better for Ultrasound: Counterintuitively, smaller input image sizes (e.g., 384x384 pixels) led to better detection performance. This is attributed to the nature of ultrasound images, where higher resolutions can exacerbate speckle noise, hindering accurate feature extraction. Downscaling helps to smooth this noise.
  • The "Optimal" Configuration is a Trade-off: The best performance wasn't achieved by simply picking the largest model and largest image. Instead, the combination of the largest model (YOLOv4-P7) with the smallest input image size (384x384) yielded the highest accuracy. This suggests that for noisy medical images, a more robust model architecture can better leverage simplified input data.
  • Real-time Performance is Achievable: The study demonstrates that the optimal configuration not only achieves high accuracy but also meets the speed requirements for real-time applications, crucial for live ultrasound guidance.

Key Results

The study found that:

  • YOLOv4-P7 with 384x384-pixel input images achieved the highest detection accuracy.
  • On the Public dataset, it reached an F1-measure of 94.7% for nerve tissues.
  • On the Original dataset, it achieved an F1-measure of 80.8% for anatomical structures (nerves and vascular tissues).
  • Larger models (P5, P6, P7) generally outperformed the smaller CSP model, showing improved Precision and reduced over-detections.
  • Smaller input image sizes consistently improved Recall across all tested models.
  • The optimal configuration (YOLOv4-P7 with 384x384 input) achieved an inference time of 5.2 milliseconds (ms), equivalent to 192.4 frames per second (fps), which is well within the range for real-time US-PNB support.

Limitations and Future Work

The authors acknowledge several limitations:

  • Dataset Size and Variability: The datasets, while useful, were relatively small and collected from a single medical facility using specific devices. This limits the generalizability of the findings. Future work should involve larger, more diverse datasets acquired with different devices and probe techniques.
  • Focus on Structures Only: The current study only detects anatomical structures (nerves and vessels). For practical US-PNB, detecting the needle is also critical. Future research should aim for concurrent detection of both anatomical structures and procedural instruments.
  • Limited DCNN Architecture Comparison: While Scaled-YOLOv4 was explored, the study did not compare its performance against other DCNN architectures or custom modules that might further enhance nerve detection. Future work should investigate different DCNN architectures and specialized modules.
  • Input Image Scaling vs. Denoising: The study focused on input image scaling to mitigate noise. Future work should compare this approach with dedicated denoising techniques (e.g., despeckle filtering) or deep learning-based denoising methods to see if they offer further improvements, or if a combination is most effective.

Summary updated

BibTeX
@misc{sugino2024investigation,
  author = {Takaaki Sugino and Shinya Onogi and Rieko Oishi and Chie Hanayama and Satoki Inoue and Shinjiro Ishida and Yuhang Yao and Nobuhiro Ogasawara and Masahiro Murakawa and Yoshikazu Nakajima},
  title = {{Investigation of Appropriate Scaling of Networks and Images for Convolutional Neural Network-Based Nerve Detection in Ultrasound-Guided Nerve Blocks}},
  year = {2024},
  doi = {10.3390/s24113696},
}

Loading related content...

Figure