1. Introduction
Breast cancer represents a significant global health challenge. Mammography is the primary imaging modality used for breast cancer screening and is associated with a significant decrease in breast cancer mortality. Ongoing technological advancements have yielded improved imaging features, facilitating the early detection of pathological signs associated with the disease [
1].
Percent density (PD) refers to the relative amount of fibroglandular tissue within the breast area compared to the overall breast area. Advanced computational methods, including deep learning algorithms, have been employed to improve the accuracy of breast density estimation from mammograms. These methods analyze digital mammograms to classify and segment dense and fatty tissues, providing a quantifiable measure of breast density. This quantification is crucial as breast density is associated with breast cancer risk [
2], and its precise assessment can enhance cancer detection and screening effectiveness [
3,
4].
One fundamental task to estimate the PD is the correct delineation of the breast tissue, excluding other areas that are not of interest that appear in the mammogram, such as background, text, pectoral muscle, and/or abdomen. Once the breast region is isolated, quantitative analysis can be performed to assess various characteristics, such as breast density. Knowledge of the breast contour also allows for further analysis of breast abnormalities such as bilateral asymmetry. Standardizing the breast segmentation process helps ensure consistency and reliability in breast tissue analysis. It is essential in large-scale screening programs where multiple radiologists or automated systems may be involved in interpreting mammograms [
5].
Previous breast delineation methods incorporate threshold-based segmentation, morphological manipulation, and a combination of Hough transforms and texture features [
6,
7]. Sansone et al. [
8] analyzed two popular packages that have been proven robust against various situations and were suitable for pectoral muscle removal. Gudhe et al. [
3] proposed a new deep learning architecture that automatically estimates the area-based breast percentage density from mammograms using a weight-adaptive multitask learning approach, which improves the baseline traditional methods. The pectoral muscle is mainly seen in medio lateral-oblique (MLO) views. For this reason, most of the available tools are aimed at the pectoral muscle exclusion in MLO views rather than in cranio-caudal (CC) views in which the muscle is sometimes present.
In our previous study [
9], we introduced a fully automated framework for dense-tissue segmentation. It included breast detection, pectoral muscle exclusion, and dense-tissue segmentation. The dense-tissue segmentation step was implemented using a Convolutional Neural Network (CNN) architecture named CM-Ynet. However, breast detection and exclusion of the pectoral muscle were carried out using traditional image processing methods. While these methods generally performed well, they encountered challenges in detecting the pectoral muscle in CC images and excluding unwanted regions, such as the abdomen. In response to these challenges, this work introduces a deep learning-based method for breast region segmentation, with a specific focus on enhancing pectoral muscle removal in CC views.
The Segment Anything Model (SAM), released in 2023 by META AI, represents a significant advancement in image segmentation. As the largest model of its kind, the SAM has been trained on billions of labeled images, covering a vast array of object types and scenarios. This extensive training enables the SAM to perform segmentation tasks on nearly any object, regardless of its complexity, without additional specific training. The SAM operates by using prompts or seed points, which can be either positive (to specify areas for segmentation) or negative (to exclude certain areas). Additionally, the SAM can utilize a bounding box approach, segmenting all objects within the specified area [
10]. While the SAM is a state-of-the-art research advancement in natural image segmentation, it does not perform satisfactorily when directly applied to medical image segmentation. Recent research suggests that this limitation primarily arises from the significant domain gap between natural images and medical images [
11,
12].
The main focus of current research is to adapt the SAM to various medical image segmentation tasks. Previous studies have explored different ways to fine-tune the SAM for specific datasets or modalities, such as retinal, abdominal, and cardiac images [
13,
14,
15,
16]. Wu et al. [
17] proposed the Medical SAM Adapter, which incorporates medical knowledge into the SAM using a simple adapter technique, and tested it on 19 segmentation tasks. More recently, a thorough study on using the SAM for medical 2D images was introduced, distinguished by its extensive data collection, detailed analysis of fine-tuning options, and comprehensive performance evaluation [
18]. However, none of these studies addressed the breast segmentation problem.
In this paper, we investigate and evaluate the SAM for breast delineation using a large-scale multi-center dataset that aims to remove all non-breast tissue in MLO and CC views, thereby providing a robust preprocessing method for further mammogram evaluations. We compare our proposed SAM-breast model with traditional and deep learning-based breast segmentation approaches, evaluating their performance on both proprietary and publicly available datasets.
3. Results
All the models described were tested on the different datasets, and a summary of the results is presented in
Table 2. Based on these results, our proposed SAM-breast model performs the best for most metrics across all datasets. The only exception is the MedSegDiff model, which shows the best distance metrics (HD and ASD). However, when considering the overall results and visual assessments, SAM-breast shows superior performance. Segmentation examples from each model and dataset are depicted in
Figure 4.
In
Table 3, we present the results for the images in CC views. This table demonstrates that all models perform exceptionally well in CC views. However, the metrics do not account for the small percentage of pectoral muscle present in the CC views. This is further illustrated in
Figure 5, where it is evident that the proposed SAM-breast model successfully excludes the pectoral muscle, even in these images.
The MLO views are significantly more complex due to the presence of the pectoral muscle. Consequently,
Table 4 displays lower metrics primarily for the thresholding method and the original SAM.
As outlined earlier, the proprietary datasets utilized in our study were sourced from various centers and acquisition devices.
Table 5 presents the results corresponding to each distinct device within the proprietary datasets. This analysis confirms that the metrics obtained are consistent across different acquisition devices.
4. Discussion
We have presented the SAM-breast model for delineating the breast area in digital mammograms. Our results demonstrate the superior performance of the proposed model compared to traditional methods and other deep learning models.
The traditional thresholding technique used in this study has proven to be successful in a large number of scenarios. However, it does not perform as expected when the image contains the abdomen or when the pectoral muscle is included in the CC views [
9]. The SAM-breast model effectively addresses this issue by excluding the pectoral muscle in both the MLO and CC views.
The effectiveness of the MedSegDiff model has been thoroughly evaluated across a broad range of medical image segmentation tasks. Specifically, it has been tested on 20 tasks involving different image modalities. This extensive testing underscores the model’s versatility and adaptability to various types of medical imaging data. It has demonstrated a significant improvement in segmentation accuracy compared to previous image segmentation models [
24]. However, in our tests, MedSegDiff did not perform as well as SAM-breast. One drawback of the MedSegDiff model is that it takes about two minutes on a single GPU to produce one prediction, while SAM-breast only takes about one second.
The original SAM can generate segmentation masks with impressive speed and quality in most natural images [
10]. However, its performance was not optimal on the different mammogram test sets. These results reinforce the necessity of adapting the SAM for specific tasks, especially for medical images.
To compare our results with previously published methodologies, we calculated the accuracy of the SAM-breast model for the Mini-MIAS and InBreast datasets. This was performed in light of the fact that accuracy is frequently employed as an evaluation metric in the majority of studies conducted on these datasets. We present a comparison with this metric in
Table 6. The most recent study that aligns with our work is by Zhou et al. [
28]. They implemented the Deeplab v3+ model, incorporating preprocessing steps such as noise suppression and contrast enhancement, which resulted in a DSC of 98.48% and IoU of 97.39% for the Mini-MIAS dataset. These values are slightly higher than those achieved by our model on the same dataset (DSC 98.07% and IoU 96.29%). However, it is important to note that their model was trained and tested on different subsets of the Mini-MIAS dataset. In contrast, our proposed model was trained on a proprietary dataset, implying that the public datasets were not seen during the training phase. This is indicative of the robust generalization performance of our model. For example, Zhou et al. also evaluated their model independently on the entire InBreast dataset (comprising 410 images), achieving a DSC of 98.48% and an IoU of 97.14%, whereas our model, tested on 200 MLO images of the InBreast dataset, achieved a DSC of 99.27% and an IoU of 98.55%. Therefore, our model consistently delivers high performance, even when analyzing images with different resolutions and originating from various vendors. We would like to highlight that our approach did not involve any image preprocessing. Instead, our focus was on extensive data augmentation during the training phase. This was achieved by simulating various contrast and intensity variations, thereby enabling us to develop a model that exhibits robust performance across a wide range of scenarios.
5. Conclusions
In this study, we presented the SAM-breast model, a specialized adaptation of the SAM, designed for the precise segmentation of the breast region in mammographic imaging. Our findings indicate that the SAM-breast model proficiently delineates the breast area, effectively excluding the pectoral muscle across both MLO and, notably, CC views where its presence is less frequent. The robustness of our model was validated across diverse datasets sourced from various acquisition devices, showcasing consistent performance throughout. The success of the SAM-breast model underscores the versatility and adaptability of the SAM framework when customized for specific medical imaging tasks, such as the segmentation of breast tissue in mammograms. This advancement holds significant promise for enhancing the accuracy and efficiency of breast cancer screening protocols.