Late Fusion-Based Deep Learning for Breast Cancer Classification in Mammography

<div>Abstract<p>Mammographic percentage of volumetric density is an important risk factor for breast cancer. Epidemiology studies historically used film images often limited to craniocaudal (CC) views to estimate area-based breast density. More recent studies using digital mammography images typically use the averaged density between craniocaudal (CC) and mediolateral oblique (MLO) view mammography for 5- and 10-year risk prediction. The performance in using either and both mammogram views has not been well-investigated. We use 3,804 full-field digital mammograms from the Joanne Knight Breast Health Cohort (294 incident cases and 657 controls), to quantity the association between volumetric percentage of density extracted from either and both mammography views and to assess the 5 and 10-year breast cancer risk prediction performance. Our results show that the association between percent volumetric density from CC, MLO, and the average between the two, retain essentially the same association with breast cancer risk. The 5- and 10-year risk prediction also shows similar prediction accuracy. Thus, one view is sufficient to assess association and predict future risk of breast cancer over a 5 or 10-year interval.</p>Prevention Relevance:<p>Expanding use of digital mammography and repeated screening provides opportunities for risk assessment. To use these images for risk estimates and guide risk management in real time requires efficient processing. Evaluating the contribution of different views to prediction performance can guide future applications for risk management in routine care.</p></div>

Handcrafted features and late fusion with deep learning for bird sound classification

In-vivo dosimetry in 3D conformal radiotherapy of breast and head and neck cancer

Microinvasive adenocarcinoma (MIA) and invasive adenocarcinoma (IAC) require distinct treatment strategies and are associated with different prognoses, underscoring the importance of accurate differentiation. This study aims to develop a predictive model that combines radiomics and deep learning to effectively distinguish between MIA and IAC. In this retrospective study, 252 pathologically confirmed cases of ground-glass nodules (GGNs) were included, with 177 allocated to the training set and 75 to the testing set. Radiomics, 2D deep learning, and 3D deep learning models were constructed based on CT images. In addition, two fusion strategies were employed to integrate these modalities: early fusion, which concatenates features from all modalities prior to classification, and late fusion, which ensembles the output probabilities of the individual models. The predictive performance of all five models was evaluated using the area under the receiver operating characteristic curve (AUC), and DeLong’s test was performed to compare differences in AUC between models. The radiomics model achieved an AUC of 0.794 (95% CI: 0.684–0.898), while the 2D and 3D deep learning models achieved AUCs of 0.754 (95% CI: 0.594–0.882) and 0.847 (95% CI: 0.724–0.945), respectively, in the testing set. Among the fusion models, the late fusion strategy demonstrated the highest predictive performance, with an AUC of 0.898 (95% CI: 0.784–0.962), outperforming the early fusion model, which achieved an AUC of 0.857 (95% CI: 0.731–0.936). Although the differences were not statistically significant, the late fusion model yielded the highest numerical values for diagnostic accuracy, sensitivity, and specificity across all models. The fusion of radiomics and deep learning features shows potential in improving the differentiation of MIA and IAC in GGNs. The late fusion strategy demonstrated promising results, warranting further validation in larger, multicenter studies.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-13447-9.

Current practice regarding reading previous films in the NHS Breast Screening Programme

Background Mammography is considered an ideal tool for breast cancer screening due to its proven effectiveness, which has led to its widespread use as the primary imaging modality for early detection. In addition, mammography is associated with a radiation dose represented as the mean glandular dose (MGD), resulting in a risk of developing breast cancer. However, efforts are being made to reduce this radiation dose. The objective of this study was to identify the MGD during screening mammography based on breast densities and investigate the correlation between MGD and BMI, breast thickness, and compression force in digital mammography. A total of 176 participants, exposed to two basic projections cranio-caudal (CC) projection and medio-lateral oblique (MLO) projection. Breast density was classified according to the guidelines by the American College of Radiology as Breast Imaging Reporting and Data System. Results Mean glandular dose (MGD) increased progressively with breast density, ranging from 8.80 ± 2.74 mGy in Category A (fatty) to 11.27 ± 3.43 mGy in Category D (extremely dense). MLO projections consistently showed higher MGD than CC views. A strong positive correlation was observed between breast thickness and MGD (r = 0.80 in left MLO). MGD also showed a moderate positive correlation with BMI (r = 0.44), while compression force had minimal impact. Conclusion Breast density, breast thickness, and BMI significantly influence the mean glandular dose in digital mammography. High breast density and thickness, along with increased BMI, are associated with elevated radiation doses, highlighting the need for personalized dose optimization to ensure effective and safe breast cancer screening.

The purpose of this study was to assess the utility of contrast-enhanced breast magnetic resonance imaging (MRI) in identifying lesions unidentified on the craniocaudal projection. The authors reviewed five patients with suspicious mammographic lesions not imaged on the craniocaudal mammogram who were referred for contrast-enhanced MRI and underwent subsequent preoperative needle localization in four of the five cases. Five patients, ages 56 to 69 years, had suspicious lesions identified on mediolateral oblique (MLO) or mediolateral (ML) projections only. Ultrasound did not identify the lesion in any of these cases. MRI identified suspicious breast lesions measuring 5 to 12 mm in size. These were located high on the chest wall or in the upper inner quadrant. Suspicious lesions seen only on the MLO or ML projections may reside high on the chest wall or in the upper inner quadrant. Lesions in these locations may be typically excluded on the craniocaudal projection during mammography. Breast MRI has the advantage of imaging the entire breast and is particularly useful for these lesions. In this series, MRI prevented delay in breast cancer diagnosis.

Breast cancer diagnosis using mammography can be significantly enhanced by integrating information from multi-source fusion and complementary imaging perspectives. In this paper, we propose a novel deep learning framework with dataset fusion for breast lesion classification. In particular, our approach fuses mammogram datasets of both Medio-Lateral-Oblique (MLO) and Cranio-Caudal (CC) views in four breast lesion categories from two data sources to ensure comprehensive lesion representation, and a combined EfficientNet-B7 and Multi-Head Attention (MHA) network is applied to classify these breast lesion categories. Therefore, the largest Region of Interest (ROI) in each image is extracted and then colorized before inserted into the network for lesion classification. Two colorized MLO and CC ROIs are separately processed by two parallel EfficientNet-B7 networks to produce the combined high-level features. The combined feature vectors from the dual EfficientNet-B7 network are then concatenated to yield a unified feature representation, which is subsequently refined using the MHA network to prioritize the most salient patterns and suppress background noise before classified using a fully connected head. To select an optimal classification model for these breast lesion categories, we performed four different models for two cases with and without an MHA network. Therefore, extensive experiments on integrated multi-source mammogram datasets demonstrate that the proposed framework outperforms conventional single-view and non-attentive models, achieving around 93.9% accuracy for breast lesion classification. The results highlight the value of dataset fusion and MHA-based feature selection in advancing computer-aided mammography.

According to a position paper of the European Commission Initiative on Breast Cancer (ECIBC), DBT is close to being introduced in European breast cancer screening programmes. Our study aimed to examine radiation dose delivered by digital breast tomosynthesis (DBT) and digital mammography (FFDM) in comparison to sole FFDM in a clinical follow-up setting and in an identical patient cohort. Retrospectively, 768 breast examinations of 96 patients were included. Patients received both DBT and FFDM between May 2015 and July 2019: (I) FFDM in cranio-caudal (CC) and DBT in mediolateral oblique (MLO) view, as well as a (II) follow-up examination with FFDM in CC and MLO view. The mean glandular dose (MGD) was determined by the mammography system according to Dance’s model. The MGD (standard deviation (SD), interquartile range (IQR)) was distributed as follows: (I) (CCFFDM+MLODBT) (a) left FFDMCC 1.40 mGy (0.36 mGy, 1.13–1.59 mGy), left DBTMLO 1.62 mGy (0.51 mGy, 1.27–1.82 mGy); (b) right FFDMCC 1.36 mGy (0.34 mGy, 1.14–1.51 mGy), right DBTMLO 1.59 mGy (0.52 mGy, 1.27–1.62 mGy). (II) (CCFFDM+MLOFFDM) (a) left FFDMCC 1.35 mGy (0.35 mGy, 1.10–1.60 mGy), left FFDMMLO 1.40 mGy (0.39 mGy, 1.12–1.59 mGy), (b) right FFDMCC 1.35 mGy (0.33 mGy, 1.12–1.48 mGy), right FFDMMLO 1.40 mGy (0.36 mGy, 1.14–1.58 mGy). MGD was significantly higher for DBT mlo views compared to FFDM (p < 0.001). Radiation dose was significantly higher for DBT in MLO views compared to FFDM. However, the MGD of DBT MLO lies below the national diagnostic reference level of 2 mGy for an FFDM view. Hence, our results support the use of either DBT or FFDM as suggested in the ECIBC’s Guidelines.

Towards personalized compression in mammography: A comparison study between pressure- and force-standardization

Mammography is one of the main diagnostic tools for the breast and is increasingly utilized for breast cancer risk assessment. Recently, parenchymal analysis has emerged as a computational tool that works by extracting imaging features from mammograms in order to infer the level of risk of a patient. In standard screening mammography, two views for each breast are obtained: the cranio-caudal (CC) and the medio-lateral oblique (MLO) views. However, to the best of our knowledge, the question of whether the choice of a given view can influence the performance of parenchymal analysis has not been researched in the literature. The aim of this work is to evaluate the utilization of different views for risk estimation based on the computerized analysis of parenchymal patterns in mammography images. We implemented a parenchymal analysis method and tested it on a sample of 228 women in a retrospective case/control study. Based on the obtained results we support the use of the CC view for parenchymal analysis since this reduces computational cost without affecting performance.

---------------------------------------------------------------------***--------------------------------------------------------------------- Abstract– Cardiovascular diseases remain one of the leading causes of mortality worldwide, making early and reliable risk prediction a critical requirement in modern healthcare systems. Traditional diagnostic approaches often rely on either clinical parameters or medical imaging in isolation, which may limit the predictiveness of the disease. This project aims to build a multimodal heart disease prediction framework that integrates clinical tabular data and echocardiography-based image information to estimate the probability of heart disease in a robust way. The tabular dataset includes demographics and clinical risk factors such as age, sex, chest pain type, blood pressure, cholesterol, electrocardiographic findings, and exercise-related parameters. Multiple classical machine learning models were trained and evaluated, including Logistic Regression, Support Vector Machine, Random Forest, Decision Tree, and XGBoost. Among all these, XGBoost achieved the best performance, with a test accuracy of 89.13% and an F1-score of 0.9029, demonstrating strong predictive capability. For the imaging modality, an echocardiography video-based model was trained using the EchoNet Dynamic dataset, where cardiac abnormality was determined based on an ejection fraction threshold of 50%. The image model achieved a ROC-AUC of 0.915, indicating excellent discriminative performance. Threshold optimization was performed to improve clinical sensitivity, increasing recall from 66.77% to 72.70% and reducing false negative cases. A late fusion approach using weighted averaging of predicted probabilities was adopted to combine both modalities, leveraging their complementary strengths. The system outputs a probabilistic risk score rather than a definitive diagnosis, making it suitable as a clinical decision support tool. Future enhancements include the integration of a RAG-based assistant for user guidance and the adoption of explainable AI techniques to improve transparency and interpretability. Key Words: Heart Disease Prediction, Multimodal Learning, Echocardiography, Machine Learning, Deep Learning, Late Fusion, Clinical Decision Support, Prediction Threshold

To evaluate the quality of breast positioning for mediolateral oblique (MLO) and craniocaudal (CC) views, a prospective study of 1,000 consecutive bilateral screening mammographic examinations was performed. Six criteria were tested, including depth of tissue seen, inferior extent of the pectoral muscle relative to the posterior nipple line, presence of fibroglandular tissue at the posterior edge of the film, and whether the nipple was in profile. Pectoral muscle was depicted to within 1 cm of the nipple line or below it on 1,612 of the 2,000 MLO mammograms (81%); all fibroglandular tissue was depicted on 1,532 MLO mammograms (77%). The depth of tissue depicted on the CC mammogram was within 1 cm greater or less than the depth on the MLO mammogram on 1,586 CC mammograms (79%); the pectoral muscle was seen on 646 CC mammograms (32%). The nipple was in profile in 1,769 MLO mammograms (88%) and 1,783 CC mammograms (89%) but not in profile in either view in 83 cases (4%). Overall improvement was seen in 400 of 587 examinations (68%) when new mammograms were compared with previous mammograms. These criteria can be used to evaluate positioning performance and for quality control.

This work presents a complete and automatic software application to aid radiologists in breast cancer diagnosis. The application is a fully automated method that performs a complete registration of magnetic resonance (MR) images and x-ray (XR) images in both directions (from MR to XR and from XR to MR) and for both x-ray mammograms, craniocaudal (CC), and mediolateral oblique (MLO). This new approximation allows radiologists to mark points in the MR images and, without any manual intervention, it provides their corresponding points in both types of XR mammograms and vice versa. The application automatically segments magnetic resonance images and x-ray images using the C-Means method and the Otsu method, respectively. It compresses the magnetic resonance images in both directions, CC and MLO, using a biomechanical model of the breast that distinguishes the specific biomechanical behavior of each one of its three tissues (skin, fat, and glandular tissue) separately. It makes a projection of both compressions and registers them with the original XR images using affine transformations and nonrigid registration methods. The application has been validated by two expert radiologists. This was carried out through a quantitative validation on 14 data sets in which the Euclidean distance between points marked by the radiologists and the corresponding points obtained by the application were measured. The results showed a mean error of 4.2 ± 1.9 mm for the MRI to CC registration, 4.8 ± 1.3 mm for the MRI to MLO registration, and 4.1 ± 1.3 mm for the CC and MLO to MRI registration. A complete software application that automatically registers XR and MR images of the breast has been implemented. The application permits radiologists to estimate the position of a lesion that is suspected of being a tumor in an imaging modality based on its position in another different modality with a clinically acceptable error. The results show that the application can accelerate the mammographic screening process for high risk populations or for dense breasts.

Cancer risk estimation in Digital Breast Tomosynthesis using GEANT4 Monte Carlo simulations and voxel phantoms