Trustworthy Breast Ultrasound Image Semantic Segmentation Based on Fuzzy Uncertainty Reduction

Abstract

Medical image semantic segmentation is essential in computer-aided diagnosis systems. It can separate tissues and lesions in the image and provide valuable information to radiologists and doctors. The breast ultrasound (BUS) images have advantages: no radiation, low cost, portable, etc. However, there are two unfavorable characteristics: (1) the dataset size is often small due to the difficulty in obtaining the ground truths, and (2) BUS images are usually in poor quality. Trustworthy BUS image segmentation is urgent in breast cancer computer-aided diagnosis systems, especially for fully understanding the BUS images and segmenting the breast anatomy, which supports breast cancer risk assessment. The main challenge for this task is uncertainty in both pixels and channels of the BUS images. In this paper, we propose a Spatial and Channel-wise Fuzzy Uncertainty Reduction Network (SCFURNet) for BUS image semantic segmentation. The proposed architecture can reduce the uncertainty in the original segmentation frameworks. We apply the proposed method to four datasets: (1) a five-category BUS image dataset with 325 images, and (2) three BUS image datasets with only tumor category (1830 images in total). The proposed approach compares state-of-the-art methods such as U-Net with VGG-16, ResNet-50/ResNet-101, Deeplab, FCN-8s, PSPNet, U-Net with information extension, attention U-Net, and U-Net with the self-attention mechanism. It achieves 2.03%, 1.84%, and 2.88% improvements in the Jaccard index on three public BUS datasets, and 6.72% improvement in the tumor category and 4.32% improvement in the overall performance on the five-category dataset compared with that of the original U-shape network with ResNet-101 since it can handle the uncertainty effectively and efficiently.

Breast tumor segmentation in ultrasound using distance-adapted fuzzy connectedness, convolutional neural network, and active contour

“…However, the experiments in Section 3.1 justify the architecture settings. Note that the mean Dice index of 0.79 obtained by the method falls within the range reported in individual benchmark collections (0.60-0.87) by most up-to-date studies from 2022-2023 [25][26][27][28][29] . Our model was trained and validated over a combined database, making it more robust and insensitive to the variability of data sources, acquisition equipment and parameters, and expert delineation specifics.…”

“…Table 3 shows our results compared with the existing solutions described in Section 1. For comparison, we selected only methods (1) described in the last two years (2022-2023) in recognized sources, (2) trained or tested using at least two of the three public US image databases used in this study [25][26][27][28][29] (Ru et al 29 employed all three databases). All methods rely on deep neural networks with architectures usually deeper and more advanced than those proposed in this paper; all apply the testing strategy within single databases.…”

“…Before the expansion of deep learning approaches, breast tumors were extracted in US scans using various canonical methods based on thresholding, watersheds, graphs, clustering, active contours, region analysis, Markov random fields, and classical neural networks (valuable surveys can be found in 20,21 ), but their efficiency was limited in high-noise images. The current state-of-the-art segmentation methods are concentrated around deep networks [22][23][24][25][26][27][28][29] . Almajalid et al 22 employed a U-Net convolutional neural network (CNN) with some pre-and postprocessing and obtained a Dice similarity coefficient (Dice index) of 0.83 over a private set of 221 images.…”

“…They reported a Dice index at 0.78 over Dataset B and 0.91 in another BUSIS database 32 . Huang et al 25 combined fuzzy logic with the CNN. In their SCFURNet model, they equipped a U-Net architecture with spatial-wise and channel-wise fuzzy blocks for uncertainty reduction, yielding a Dice index of 0.81 in both BUSI and Dataset B (plus 0.93 in BUSIS database and 0.82 in a multiclass task over a private collection).…”

Incorporating Tumor Edge Information for Fine-Grained BI-RADS Classification of Breast Ultrasound Images

“…The classification accuracy drops when applying the vertical flip technique, possibly due to the specific characteristics of BUS images. In BUS images, there are five layers from top to bottom, including the pre-fat background area, fat layer, mammary layer, muscle layer, and retro-muscle layer, and breast tumors are generally located in the mammary layer [48], [49]. Therefore, changing the vertical orientation of BUS images may disrupt the network's learning of specific patterns from BUS images.…”