HIFUNet: Multi-Class Segmentation of Uterine Regions From MR Images Using Global Convolutional Networks for HIFU Surgery Planning

Abstract

Accurate segmentation of uterus, uterine fibroids, and spine from MR images is crucial for high intensity focused ultrasound (HIFU) therapy but remains still difficult to achieve because of 1) the large shape and size variations among individuals, 2) the low contrast between adjacent organs and tissues, and 3) the unknown number of uterine fibroids. To tackle this problem, in this paper, we propose a large kernel Encoder-Decoder Network based on a 2D segmentation model. The use of this large kernel can capture multi-scale contexts by enlarging the valid receptive field. In addition, a deep multiple atrous convolution block is also employed to enlarge the receptive field and extract denser feature maps. Our approach is compared to both conventional and other deep learning methods and the experimental results conducted on a large dataset show its effectiveness.

Deep learning enables automated MRI-based estimation of uterine volume also in patients with uterine fibroids undergoing high-intensity focused ultrasound therapy

“…Precise segmentation was reported to be difficult at the margins for patients with many fibroids, resulting in a Dice score for uterine segmentation of 82.37%. The authors suggest that direct 3D segmentation may lead to higher segmentation performance [ 32 ]. Because post-HIFU image data were not included in this study, information on the applicability of this method for accurate assessment of treatment-related volume changes over time after HIFU ablation is missing.…”

“…It should be noted that the precise contours, especially of small fibroids, are often difficult to delineate from adjacent uterine tissue, even for human readers. This may also contribute to lower accuracy of segmentation of individual fibroids compared to uterine segmentation [ 32 , 37 ]. From a clinical perspective, segmentation of the entire uterus may already be a relevant measure for therapy response assessment.…”

“…Very recently, various neural networks have been proposed for uterine segmentation in MRI and ultrasound imaging, where most of them are also based on the U-Net architecture [29][30][31][32]. However, none of these has presented a suitable method for accurate automatic volumetry of the uterus, especially when the evaluation of longitudinal data is required to assess treatment response, such as in patients with uterine fibroids undergoing HIFU therapy.…”

CTANet: Confidence-Based Threshold Adaption Network for Semi-Supervised Segmentation of Uterine Regions from MR Images for HIFU Treatment

“…We compared our CTANet with four SOTA semi-supervised learning approaches, including ASDNet [38], Latent Mixup [43], and Cross-Consistency Training (CCT) [37], this for 3 different labeled/unlabeled data ratios. Besides, we added two fully-supervised methods: the classical Vanilla U-Net [6] and HI-FUNet [21] with the whole set of labeled data as the performance reference. All the experiments were conducted in a fair way with the same training, test data, and network hyperparameters.…”

“…Kurata et al [20] used an optimized U-Net to segment the uterus on T2-weighted MR images. Zhang et al [21] proposed a large kernel Encoder-Decoder Network based on a 2D segmentation model named HIFUNet to segment all uterine regions from MR images. Ning et al [22] developed a multistage segmentation network to segment the fibroids.…”