Deep learning‐based reconstruction can improve canine thoracolumbar magnetic resonance image quality and reduce slice thickness

Abstract

In veterinary practice, thin‐sliced thoracolumbar MRI is useful in detecting small lesions, especially in small‐breed dogs. However, it is challenging due to the partial volume averaging effect and increase in scan time. Currently, deep learning‐based reconstruction (DLR), a part of artificial intelligence, has been applied in diagnostic imaging. We hypothesized that the diagnostic performance of thin‐slice thoracolumbar MRI with DLR would be superior to conventional MRI. This prospective, method comparison study aimed to determine the adequate slice thickness of a deep learning model for thin‐slice thoracolumbar MRI. Sagittal and transverse T2‐weighted MRI at the thoracolumbar region were performed on 12 clinically healthy beagle dogs; the images obtained were categorized into five groups according to slice thickness: conventional thickness of 3 mm (3CON) and thicknesses of 3, 2, 1.5, and 1 mm with DLR (3DLR, 2DLR, 1.5DLR, and 1DLR, respectively). Quantitative analysis was performed using signal‐to‐noise ratio (SNR) and contrast‐to‐noise ratio. Qualitative analysis involved the evaluation of perceived SNR, structural visibility, and overall image quality using a four‐point scale. Moreover, nerve root visibility was evaluated using transverse images. Quantitative and qualitative values were compared among the five groups. Compared with the 3CON group, the 3DLR, 2DLR, and 1.5DLR groups exhibited significantly higher quantitative and qualitative values. Nerve root visibility was significantly higher in 2DLR, 1.5DLR, and 1DLR images than in 3DLR and 3CON images. Compared with conventional MRI, DLR reduced the slice thickness by up to one‐half and improved image quality in this sample of clinically healthy beagles.

Exploring Deep Learning Strategies for Intervertebral Disc Herniation Detection on Veterinary MRI

“…Such integration has the potential to signi cantly improve diagnostic accuracy and inform more precise treatment strategies. From another perspective, optimizing imaging protocols and enhancing image reconstruction quality 15 The spine localization module led to higher detection accuracy at nearly all recall levels. Detection accuracy on the COCO dataset versus on the canine IVDH dataset for various models.…”

“…While arti cial intelligence (AI) or deep learning research has advanced IVDH detection in humans [10][11][12][13][14] , the anatomical differences between humans and animals, especially the smaller size of animal intervertebral discs, create unique challenges in adapting these methods for veterinary applications. Moreover, while a few studies 15,16 have applied AI techniques to canine IVDH, their focus has predominantly been on image quality improvement 15 or image-level disease classi cation 16 . The localization of IVDH lesions on segment level, as an arguably more challenging yet crucial task, remains largely unexplored yet.…”