Breast Segmentation Research Articles

VMC‐UNet: A Vision Mamba‐CNN U‐Net for Tumor Segmentation in Breast Ultrasound Image

ABSTRACTBreast cancer remains one of the most significant health threats to women, making precise segmentation of target tumors critical for early clinical intervention and postoperative monitoring. While numerous convolutional neural networks (CNNs) and vision transformers have been developed to segment breast tumors from ultrasound images, both architectures encounter difficulties in effectively modeling long‐range dependencies, which are essential for accurate segmentation. Drawing inspiration from the Mamba architecture, we introduce the Vision Mamba‐CNN U‐Net (VMC‐UNet) for breast tumor segmentation. This innovative hybrid framework merges the long‐range dependency modeling capabilities of Mamba with the detailed local representation power of CNNs. A key feature of our approach is the implementation of a residual connection method within the U‐Net architecture, utilizing the visual state space (VSS) module to extract long‐range dependency features from convolutional feature maps effectively. Additionally, to better integrate texture and structural features, we have designed a bilinear multi‐scale attention module (BMSA), which significantly enhances the network's ability to capture and utilize intricate feature details across multiple scales. Extensive experiments conducted on three public datasets demonstrate that the proposed VMC‐UNet surpasses other state‐of‐the‐art methods in breast tumor segmentation, achieving Dice coefficients of 81.52% for BUSI, 88.00% for BUS, and 88.96% for STU. The source code is accessible at https://github.com/windywindyw/VMC‐UNet.

Progressive Dual Priori Network for Generalized Breast Tumor Segmentation.

To promote the generalization ability of breast tumor segmentation models, as well as to improve the segmentation performance for breast tumors with smaller size, low-contrast and irregular shape, we propose a progressive dual priori network (PDPNet) to segment breast tumors from dynamic enhanced magnetic resonance images (DCE-MRI) acquired at different centers. The PDPNet first cropped tumor regions with a coarse-segmentation based localization module, then the breast tumor mask was progressively refined by using the weak semantic priori and cross-scale correlation prior knowledge. To validate the effectiveness of PDPNet, we compared it with several state-of-the-art methods on multi-center datasets. The results showed that, comparing against the suboptimal method, the DSC and HD95 of PDPNet were improved at least by 5.13% and 7.58% respectively on multi-center test sets. In addition, through ablations, we demonstrated that the proposed localization module can decrease the influence of normal tissues and therefore improve the generalization ability of the model. The weak semantic priors allow focusing on tumor regions to avoid missing small tumors and low-contrast tumors. The cross-scale correlation priors are beneficial for promoting the shape-aware ability for irregular tumors. Thus integrating them in a unified framework improved the multi-center breast tumor segmentation performance.

Breast segmentation in infrared thermography from characteristical inframammary shape.

Infrared thermography is gaining relevance in breast cancer assessment. For this purpose, breast segmentation in thermograms is an important task for performing automatic image analysis and detecting possible temperature changes that indicate the presence of malignancy. However, it is not a simple task since the breast limit borders, especially the top borders, often have low contrast, making it difficult to isolate the breast area. Several algorithms have been proposed for breast segmentation, but these highly depend on the contrast at the lower breast borders and on filtering algorithms to remove false edges. This work focuses on taking advantage of the distinctive inframammary shape to simplify the definition of the lower breast border, regardless of the contrast level, which indeed also provides a strong anatomical reference to support the definition of the poorly marked upper boundary of the breasts, which has been one of the major challenges in the literature. In order to demonstrate viability of the proposed technique for an automatic breast segmentation, we applied it to a database with 180 thermograms and compared their results with those reported by others in the literature. We found that our approach achieved a high performance, in terms of Intersection over Union of 0.934, even higher than that reported by artificial intelligence algorithms. The performance is invariant to breast sizes and thermal contrast of the images.

A deep learning-based method for the detection and segmentation of breast masses in ultrasound images

Objective. Automated detection and segmentation of breast masses in ultrasound images are critical for breast cancer diagnosis, but remain challenging due to limited image quality and complex breast tissues. This study aims to develop a deep learning-based method that enables accurate breast mass detection and segmentation in ultrasound images. Approach. A novel convolutional neural network-based framework that combines the You Only Look Once (YOLO) v5 network and the Global-Local (GOLO) strategy was developed. First, YOLOv5 was applied to locate the mass regions of interest (ROIs). Second, a Global Local-Connected Multi-Scale Selection (GOLO-CMSS) network was developed to segment the masses. The GOLO-CMSS operated on both the entire images globally and mass ROIs locally, and then integrated the two branches for a final segmentation output. Particularly, in global branch, CMSS applied Multi-Scale Selection (MSS) modules to automatically adjust the receptive fields, and Multi-Input (MLI) modules to enable fusion of shallow and deep features at different resolutions. The USTC dataset containing 28 477 breast ultrasound images was collected for training and test. The proposed method was also tested on three public datasets, UDIAT, BUSI and TUH. The segmentation performance of GOLO-CMSS was compared with other networks and three experienced radiologists. Main results. YOLOv5 outperformed other detection models with average precisions of 99.41%, 95.15%, 93.69% and 96.42% on the USTC, UDIAT, BUSI and TUH datasets, respectively. The proposed GOLO-CMSS showed superior segmentation performance over other state-of-the-art networks, with Dice similarity coefficients (DSCs) of 93.19%, 88.56%, 87.58% and 90.37% on the USTC, UDIAT, BUSI and TUH datasets, respectively. The mean DSC between GOLO-CMSS and each radiologist was significantly better than that between radiologists (p < 0.001). Significance. Our proposed method can accurately detect and segment breast masses with a decent performance comparable to radiologists, highlighting its great potential for clinical implementation in breast ultrasound examination.

Deep learning methods for cancer recognition

Modern medical technology has been advancing continuously, and deep learning, a potent machine learning technique, demonstrates significant potential in the domain of cancer identification and diagnosis. This paper through methods of lieterature review, examines the distinctive abilities and advancements of deep learning in the recognition, classification, and segmentation of breast and lung cancer images. Deep learning has proven to be very suitable for medical image analysis and has the ability to aid and even autonomously make decisions in many stages of cancer diagnosis, including history taking, imaging, and biopsy. Furthermore, it possesses the capability to precisely identify, categorize, and divide photos pertaining to breast and lung cancer. Additionally, it has the capacity to identify and predict tumors based on gene expression profiles, so enhancing the efficacy of contemporary medicine in the identification and prevention of cancer. These findings have significant implications for cancer detection and treatment in contemporary healthcare, offering clinicians a potent tool to enhance diagnostic precision and treatment effectiveness.

MRI-based Deep Learning Models for Preoperative Breast Volume and Density Assessment Assisting Breast Reconstruction.

The volume of the implant is the most critical element of breast reconstruction, so it is necessary to accurately assess the preoperative volume of the healthy and affected breasts and select the appropriate implant for placement. Accurate and automated methods for quantitative assessment of breast volume can optimize breast reconstruction surgery and assist physicians in clinical decision making. The aim of this study was to develop an artificial intelligence model for automated segmentation of the breast and measurement of volume. A total of 249 subjects undergoing breast reconstruction surgery were enrolled in this study. Subjects underwent preoperative breast MRI, and the breast region manually outlined by the imaging physician served as the gold standard for volume measurement by the automated segmentation model. In this study, we developed three automated algorithms for automatic segmentation of breast regions, including a simple alignment model, an alignment dynamic encoding model, and a deep learning model. The volumetric agreement between the three automated segmentation algorithms and the breast regions manually segmented by imaging physicians was evaluated by calculating the mean square error (MSE) and intragroup correlation coefficient (ICC), and the reproducibility of the automated segmentation of the breast regions was assessed by the test-retest step. The three breast automated segmentation models developed in this study (simple registration model, dynamic programming model, and deep learning model) showed strong ICC with manual segmentation of the breast region, with MSEs of 1.124, 0.693, and 0.781, and ICCs of 0.975 (95% CI, 0.869-0.991), 0.986 (95% CI, 0.967-0.996), and 0.983 (95% CI, 0.961-0.992), respectively. Regarding the test-retest results of breast volume, the dynamic programming model performed the best with an MSE of 0.370 and an ICC of 0.993 (95% CI, 0.982-0.997), followed by the deep learning algorithm with an MSE of 0.741 and an ICC of 0.983 (95% CI, 0.956-0.993), and the simple registration algorithm with an MSE of 0.763 and an ICC of 0.982 (95% CI, 0.949-0.993). The reproducibility of the breast region segmented by the three automated algorithms was higher than that of manual segmentation by different radiologists. The three automated breast segmentation algorithms developed in this study generate accurate and reliable breast regions, enable highly reproducible breast region segmentation and automated volume measurements, and provide a valuable tool for surgical selection of appropriate prostheses. This journal requires that authors assign a level of evidence to each submission to which Evidence-Based Medicine rankings are applicable. This excludes Review Articles, Book Reviews, and manuscripts that concern Basic Science, Animal Studies, Cadaver Studies, and Experimental Studies. For a full description of these Evidence-Based Medicine ratings, please refer to the Table of Contents or the online Instructions to Authors www.springer.com/00266 .

MurSS: A Multi-Resolution Selective Segmentation Model for Breast Cancer.

Accurately segmenting cancer lesions is essential for effective personalized treatment and enhanced patient outcomes. We propose a multi-resolution selective segmentation (MurSS) model to accurately segment breast cancer lesions from hematoxylin and eosin (H&E) stained whole-slide images (WSIs). We used The Cancer Genome Atlas breast invasive carcinoma (BRCA) public dataset for training and validation. We used the Korea University Medical Center, Guro Hospital, BRCA dataset for the final test evaluation. MurSS utilizes both low- and high-resolution patches to leverage multi-resolution features using adaptive instance normalization. This enhances segmentation performance while employing a selective segmentation method to automatically reject ambiguous tissue regions, ensuring stable training. MurSS rejects 5% of WSI regions and achieves a pixel-level accuracy of 96.88% (95% confidence interval (CI): 95.97-97.62%) and mean Intersection over Union of 0.7283 (95% CI: 0.6865-0.7640). In our study, MurSS exhibits superior performance over other deep learning models, showcasing its ability to reject ambiguous areas identified by expert annotations while using multi-resolution inputs.

Computerized assessment of background parenchymal enhancement on breast dynamic contrast-enhanced-MRI including electronic lesion removal.

Current clinical assessment qualitatively describes background parenchymal enhancement (BPE) as minimal, mild, moderate, or marked based on the visually perceived volume and intensity of enhancement in normal fibroglandular breast tissue in dynamic contrast-enhanced (DCE)-MRI. Tumor enhancement may be included within the visual assessment of BPE, thus inflating BPE estimation due to angiogenesis within the tumor. Using a dataset of 426 MRIs, we developed an automated method to segment breasts, electronically remove lesions, and calculate scores to estimate BPE levels. A U-Net was trained for breast segmentation from DCE-MRI maximum intensity projection (MIP) images. Fuzzy -means clustering was used to segment lesions; the lesion volume was removed prior to creating projections. U-Net outputs were applied to create projection images of both, affected, and unaffected breasts before and after lesion removal. BPE scores were calculated from various projection images, including MIPs or average intensity projections of first- or second postcontrast subtraction MRIs, to evaluate the effect of varying image parameters on automatic BPE assessment. Receiver operating characteristic analysis was performed to determine the predictive value of computed scores in BPE level classification tasks relative to radiologist ratings. Statistically significant trends were found between radiologist BPE ratings and calculated BPE scores for all breast regions (Kendall correlation, ). Scores from all breast regions performed significantly better than guessing ( from the -test). Results failed to show a statistically significant difference in performance with and without lesion removal. BPE scores of the affected breast in the second postcontrast subtraction MIP after lesion removal performed statistically greater than random guessing across various viewing projections and DCE time points. Results demonstrate the potential for automatic BPE scoring to serve as a quantitative value for objective BPE level classification from breast DCE-MR without the influence of lesion enhancement.

1886: Validation of Deep Learning based CTV breast segmentation model on a large multicentric dataset

1886: Validation of Deep Learning based CTV breast segmentation model on a large multicentric dataset

Quantitative assessment of background parenchymal enhancement is associated with lifetime breast cancer risk in screening MRI

ObjectivesTo compare the quantitative background parenchymal enhancement (BPE) in women with different lifetime risks and BRCA mutation status of breast cancer using screening MRI.Materials and methodsThis study included screening MRI of 535 women divided into three groups based on lifetime risk: nonhigh-risk women, high-risk women without BRCA mutation, and BRCA1/2 mutation carriers. Six quantitative BPE measurements, including percent enhancement (PE) and signal enhancement ratio (SER), were calculated on DCE-MRI after segmentation of the whole breast and fibroglandular tissue (FGT). The associations between lifetime risk factors and BPE were analyzed via linear regression analysis. We adjusted for risk factors influencing BPE using propensity score matching (PSM) and compared the BPE between different groups. A two-sided Mann–Whitney U-test was used to compare the BPE with a threshold of 0.1 for multiple testing issue-adjusted p values.ResultsAge, BMI, menopausal status, and FGT level were significantly correlated with quantitative BPE based on the univariate and multivariable linear regression analyses. After adjusting for age, BMI, menopausal status, hormonal treatment history, and FGT level using PSM, significant differences were observed between high-risk non-BRCA and BRCA groups in PEFGT (11.5 vs. 8.0%, adjusted p = 0.018) and SERFGT (7.2 vs. 9.3%, adjusted p = 0.066).ConclusionQuantitative BPE varies in women with different lifetime breast cancer risks and BRCA mutation status. These differences may be due to the influence of multiple lifetime risk factors. Quantitative BPE differences remained between groups with and without BRCA mutations after adjusting for known risk factors associated with BPE.Clinical relevance statementBRCA germline mutations may be associated with quantitative background parenchymal enhancement, excluding the effects of known confounding factors. This finding can provide potential insights into the cancer pathophysiological mechanisms behind lifetime risk models.Key PointsExpanding understanding of breast cancer pathophysiology allows for improved risk stratification and optimized screening protocols.Quantitative BPE is significantly associated with lifetime risk factors and differs between BRCA mutation carriers and noncarriers.This research offers a possible understanding of the physiological mechanisms underlying quantitative BPE and BRCA germline mutations.

A publicly available deep learning model and dataset for segmentation of breast, fibroglandular tissue, and vessels in breast MRI

Breast density, or the amount of fibroglandular tissue (FGT) relative to the overall breast volume, increases the risk of developing breast cancer. Although previous studies have utilized deep learning to assess breast density, the limited public availability of data and quantitative tools hinders the development of better assessment tools. Our objective was to (1) create and share a large dataset of pixel-wise annotations according to well-defined criteria, and (2) develop, evaluate, and share an automated segmentation method for breast, FGT, and blood vessels using convolutional neural networks. We used the Duke Breast Cancer MRI dataset to randomly select 100 MRI studies and manually annotated the breast, FGT, and blood vessels for each study. Model performance was evaluated using the Dice similarity coefficient (DSC). The model achieved DSC values of 0.92 for breast, 0.86 for FGT, and 0.65 for blood vessels on the test set. The correlation between our model’s predicted breast density and the manually generated masks was 0.95. The correlation between the predicted breast density and qualitative radiologist assessment was 0.75. Our automated models can accurately segment breast, FGT, and blood vessels using pre-contrast breast MRI data. The data and the models were made publicly available.

Artificial Intelligence - A Primer for Diagnosis and Interpretation of Breast Cancer

Breast Cancer (BC) is a major universal health problem. Early detection and precise diagnosis are vital for enlightening outcomes. Artificial Intelligence (AI) technologies can potentially revolutionize the field of BC by providing quantitative representations of medical images to assist in segmentation, diagnosis, and prognosis. AI can improve image quality, detect and segment breast lesions, classify cancer and predict its behavior, and integrate data from multiple sources to predict clinical outcomes. It can lead to more personalized and effective treatment for BC patients. Challenges faced by AI in real-life solicitations include data curation, model interpretability, and run-through guidelines. However, the clinical implementation of AI is expected to deliver vital guidance for patient-tailored management. BC is a major global health problem; early detection and treatment are crucial for improving outcomes. Imaging detection is a key screening, diagnosis, and treatment effectiveness assessment tool. However, the irresistible number of images creates a heavy capacity for radiologists and delays reporting. AI has the potential to revolutionize BC imaging by improving efficiency and accuracy. AI can recognize, segment, and diagnose tumor lesions automatically and analyze tumor images on a molecular level. It could lead to more personalized treatment strategies. However, AI-assisted imaging diagnosis is still in its early stages of development, and more research is needed to validate its clinical effectiveness. Therefore, AI is a promising new technology that has the potential to progress the diagnosis and treatment of BC, and AI-assisted imaging diagnosis is a promising new technology for improving the early detection and diagnosis of BC. More research is needed to bring this technology to clinical practice.

Prospective deployment of an automated implementation solution for artificial intelligence translation to clinical radiation oncology.

Artificial intelligence (AI)-based technologies embody countless solutions in radiation oncology, yet translation of AI-assisted software tools to actual clinical environments remains unrealized. We present the Deep Learning On-Demand Assistant (DL-ODA), a fully automated, end-to-end clinical platform that enables AI interventions for any disease site featuring an automated model-training pipeline, auto-segmentations, and QA reporting. We developed, tested, and prospectively deployed the DL-ODA system at a large university affiliated hospital center. Medical professionals activate the DL-ODA via two pathways (1): On-Demand, used for immediate AI decision support for a patient-specific treatment plan, and (2) Ambient, in which QA is provided for all daily radiotherapy (RT) plans by comparing DL segmentations with manual delineations and calculating the dosimetric impact. To demonstrate the implementation of a new anatomy segmentation, we used the model-training pipeline to generate a breast segmentation model based on a large clinical dataset. Additionally, the contour QA functionality of existing models was assessed using a retrospective cohort of 3,399 lung and 885 spine RT cases. Ambient QA was performed for various disease sites including spine RT and heart for dosimetric sparing. Successful training of the breast model was completed in less than a day and resulted in clinically viable whole breast contours. For the retrospective analysis, we evaluated manual-versus-AI similarity for the ten most common structures. The DL-ODA detected high similarities in heart, lung, liver, and kidney delineations but lower for esophagus, trachea, stomach, and small bowel due largely to incomplete manual contouring. The deployed Ambient QAs for heart and spine sites have prospectively processed over 2,500 cases and 230 cases over 9 months and 5 months, respectively, automatically alerting the RT personnel. The DL-ODA capabilities in providing universal AI interventions were demonstrated for On-Demand contour QA, DL segmentations, and automated model training, and confirmed successful integration of the system into a large academic radiotherapy department. The novelty of deploying the DL-ODA as a multi-modal, fully automated end-to-end AI clinical implementation solution marks a significant step towards a generalizable framework that leverages AI to improve the efficiency and reliability of RT systems.

Mass segmentation in automated breast ultrasound using an enhanced attentive UNet

Breast cancer is the most diagnosed cancer among women. Early detection of breast cancer through screening high risk population significantly reduces the mortality rate. To this end, 3-D Automated Breast Ultrasound (ABUS) is a favorable modality due to its complementary information in comparison with other modalities. Different characteristics of breast masses and the large number of ABUS slices result in time consuming and exhaustive evaluations by radiologists. This study aims to develop a fully-automatic segmentation system that acts as a second interpreter to facilitate the interpretation process. Specifically, a bidirectional convolutional long short-term memory (Bi-ConvLSTM) network is adopted for tumor detection. Afterwards, a 2-D attentive UNet takes an ABUS slice and its approximate saliency map to precisely segment breast masses. The attention blocks of the proposed network use this map to direct the network’s attention to regions with high saliency values. We proposed two enhanced versions of the base attention module with superior performance and less complexity. To increase the robustness against the mass size variations, we stacked several side outputs from different stages of the network and control their contribution to final results by a Convolutional Block Attention Module (CBAM). The effectiveness of the proposed modifications is verified with step-by-step evaluations, and our optimal model achieves a Dice similarity index of 85.82% on a dataset of 60 volumes.

An improved semantic segmentation for breast lesion from dynamic contrast enhanced MRI images using deep learning

AbstractThe World Health Organization (WHO) reports that approximately 2.3 million breast cancer cases are diagnosed each year. Early detection is key to tackling this issue, and Dynamic Contrast Enhanced Magnetic Resonance Imaging (DCE‐MRI) is a preferred method for detecting tumors. Convolutional Neural Networks (CNNs) can accurately segment images without human assistance. The objective of this study is to develop a computer‐aided diagnosis system that can segment breast lesions from DCE‐MRI images. A 92‐layer deep CNN, called DCNN‐92, and a 94‐layer deep CNN, called DCNN‐94, have been designed to identify lesions. The proposed methods have been validated using images from The Cancer Image Archive (TCIA) database. The proposed DCNN‐92 model segments the tumor pixels effectively, but it exhibits some misclassifications where certain background pixels are incorrectly labeled as tumor and tumor pixels are identified as background. To segment the tumor pixels more accurately, two grouped convolution layers are added to the DCNN‐92 model. The model with 94 layers, that is, DCNN‐94, segments most of the tumor pixels correctly, thereby enhancing the segmentation performance. When compared to DCNN‐92, the DCNN‐94 model exhibits enhanced performance across standard metrics such as sensitivity, dice coefficient, Jaccard coefficient, and area under the curve (AUC). It was found that the training time for DCNN‐94 is shorter. The DCNN‐94 model with dilation factor and group convolution is concluded to be an effective method for lesion segmentation from breast DCE‐MRI images compared to existing methods.

BCUIS‐Net: A breast cancer ultrasound image segmentation network via boundary‐aware and shape feature fusion

AbstractBreast cancer is a highly lethal disease with the highest mortality rate among women worldwide. Breast tumor segmentation from ultrasound images plays a critical role in enabling early detection, leading to a reduction in mortality rates. However, the challenge of ultrasound breast cancer segmentation arises from factors such as indistinct lesion boundaries, noise artifacts, and inhomogeneous intensity distribution within the lesion region. To address the bottlenecks, a novel boundary‐aware shape feature fusion network (BCUIS‐Net) is proposed to segment breast lesion in ultrasound images. Firstly, a boundary‐aware module (BAM) is put forward to accurately localize the ambiguous tumor regions and boundaries by embedding the horizontal and vertical position information into the channel attention. Subsequently, a shape feature fusion (SFF) module is presented to fuse shape features and segmentation features, in order to adaptively extract their complementary features by aggregating contextual information in an attention module. Specifically, the different levels of features from the encoder are up‐sampled to the original image size and fed into the BAM to predict the boundary map. The boundary and decoder‐generated feature maps are thereafter fused by the SFF module to exploit the complementarity between them to correct errors in segmentation and shape features, effectively eliminating false detections and noise in the features to achieve accurate segmentation of pathological regions. Finally, the shape fusion loss is derived from a combination of the binary cross‐entropy loss and the distance map loss to intelligently penalize incorrect predictions and thus improve the attention to boundary locations. The performance of the network is evaluated in two public breast ultrasound datasets. Experimental results verify that the proposed method obtains superior segmentation results and outperforms the most recent state‐of‐the‐art, in which IOU is increased by 2.15% and 2.59% on UDIAT and BUSI, respectively.

Pathologic complete response prediction in breast cancer lesion segmentation and neoadjuvant therapy.

Predicting whether axillary lymph nodes could achieve pathologic Complete Response (pCR) after breast cancer patients receive neoadjuvant chemotherapy helps make a quick follow-up treatment plan. This paper presents a novel method to achieve this prediction with the most effective medical imaging method, Dynamic Contrast-enhanced Magnetic Resonance Imaging (DCE-MRI). In order to get an accurate prediction, we first proposed a two-step lesion segmentation method to extract the breast cancer lesion region from DCE-MRI images. With the segmented breast cancer lesion region, we then used a multi-modal fusion model to predict the probability of axillary lymph nodes achieving pCR. We collected 361 breast cancer samples from two hospitals to train and test the proposed segmentation model and the multi-modal fusion model. Both segmentation and prediction models obtained high accuracy. The results show that our method is effective in both the segmentation task and the pCR prediction task. It suggests that the presented methods, especially the multi-modal fusion model, can be used for the prediction of treatment response in breast cancer, given data from noninvasive methods only.

Saliency of breast lesions in breast cancer detection using artificial intelligence

The analysis of mammograms using artificial intelligence (AI) has shown great potential for assisting breast cancer screening. We use saliency maps to study the role of breast lesions in the decision-making process of AI systems for breast cancer detection in screening mammograms. We retrospectively collected mammograms from 191 women with screen-detected breast cancer and 191 healthy controls matched by age and mammographic system. Two radiologists manually segmented the breast lesions in the mammograms from CC and MLO views. We estimated the detection performance of four deep learning-based AI systems using the area under the ROC curve (AUC) with a 95% confidence interval (CI). We used automatic thresholding on saliency maps from the AI systems to identify the areas of interest on the mammograms. Finally, we measured the overlap between these areas of interest and the segmented breast lesions using Dice’s similarity coefficient (DSC). The detection performance of the AI systems ranged from low to moderate (AUCs from 0.525 to 0.694). The overlap between the areas of interest and the breast lesions was low for all the studied methods (median DSC from 4.2% to 38.0%). The AI system with the highest cancer detection performance (AUC = 0.694, CI 0.662–0.726) showed the lowest overlap (DSC = 4.2%) with breast lesions. The areas of interest found by saliency analysis of the AI systems showed poor overlap with breast lesions. These results suggest that AI systems with the highest performance do not solely rely on localized breast lesions for their decision-making in cancer detection; rather, they incorporate information from large image regions. This work contributes to the understanding of the role of breast lesions in cancer detection using AI.

Assessment of deep learning-based auto-contouring on interobserver consistency in target volume and organs-at-risk delineation for breast cancer: Implications for RTQA program in a multi-institutional study

PurposeTo quantify interobserver variation (IOV) in target volume and organs-at-risk (OAR) contouring across 31 institutions in breast cancer cases and to explore the clinical utility of deep learning (DL)-based auto-contouring in reducing potential IOV. Methods and materialsIn phase 1, two breast cancer cases were randomly selected and distributed to multiple institutions for contouring six clinical target volumes (CTVs) and eight OAR. In Phase 2, auto-contour sets were generated using a previously published DL Breast segmentation model and were made available for all participants. The difference in IOV of submitted contours in phases 1 and 2 was investigated quantitatively using the Dice similarity coefficient (DSC) and Hausdorff distance (HD). The qualitative analysis involved using contour heat maps to visualize the extent and location of these variations and the required modification. ResultsOver 800 pairwise comparisons were analysed for each structure in each case. Quantitative phase 2 metrics showed significant improvement in the mean DSC (from 0.69 to 0.77) and HD (from 34.9 to 17.9 mm). Quantitative analysis showed increased interobserver agreement in phase 2, specifically for CTV structures (5–19 %), leading to fewer manual adjustments. Underlying IOV differences causes were reported using a questionnaire and hierarchical clustering analysis based on the volume of CTVs. ConclusionDL-based auto-contours improved the contour agreement for OARs and CTVs significantly, both qualitatively and quantitatively, suggesting its potential role in minimizing radiation therapy protocol deviation.

Multimodal image feature fusion for improving medical ultrasound image segmentation

Ultrasound imaging is the most commonly used technique to detect abnormalities of the thyroid and diagnose abnormal breasts. However, due to the inherent characteristics of ultrasound imaging, the contrast of ultrasound images is relatively low, and the tissue boundaries in the image are blurry. This makes it a challenging task to segment thyroid or breast lesions from ultrasound images. In this work, we explore the role of multimodal features at different levels of deep neural network models and their fusion mechanisms for multimodal images generated by ultrasound radio frequency (RF) data. The ultrasound RF data contain information related to the arrangement of scatterers in tissues and organs, which is only partially available in B-mode images. To solve this problem, we use ultrasound RF data to generate parametric images, including Nakagami and entropy images, to illustrate the abundant backscattered statistics contained in RF data. A channel-aware fusion module is proposed to adaptively fuse the features from images of different modalities. Experimental results show that the abundant information contained in RF data can be helpful to identify thyroid glands or breast lesions with low contrast and fuzzy boundaries in B-mode ultrasound images and improve the segmentation performance.