CORTEXA
← Browse
crossrefBioengineering2026-04-17Cited by 0

Clinically Robust Deep Learning for Contrast-Enhanced Mammography: Multicenter Evaluation Across Convolutional Neural Network Architectures

Roberta Fusco, Vincenza Granata, Paolo Vallone, Teresa Petrosino, Maria Daniela Iasevoli, Roberta Galdiero, Mauro Mattace Raso, Davide Pupo, Filippo Tovecci, Annamaria Porto, Gerardo Ferrara, Modesta Longobucco, Giulia Capuano, Roberto Morcavallo, Caterina Todisco, Fabiana Antenucci, Mario Sansone, Mimma Castaldo, Daniele La Forgia, Antonella Petrillo

Background: This study investigates the impact of anatomically constrained preprocessing and deep learning architecture selection on benign versus malignant breast lesion classification in contrast-enhanced mammography (CEM), with the goal of improving robustness and clinical reliability across heterogeneous data sources. Methods: In this retrospective multicenter study, CEM images from 300 patients (314 lesions) were combined with 1003 publicly available CEM images, yielding a total of 1120 breast cases. Automatic breast segmentation was performed using the LIBRA framework to generate breast-mask images. Eleven deep learning models, including classical convolutional neural networks, attention-based networks, hybrid convolutional neural networks (CNNs), Transformer architectures, and mammography-specific models, were trained and evaluated using both original DICOM images and breast-mask inputs. Performance was assessed using accuracy, balanced accuracy, sensitivity, specificity, AUROC, and AUPRC on cross-validation and independent test sets. Hyperparameter optimization was conducted for the best-performing architecture. Results: Models trained on breast-mask images consistently outperformed those trained on original DICOM images across all architectures and metrics, with AUROC improvements ranging from +0.06 to +0.21. Among all models, ResNet50 trained on breast-mask images achieved the best performance (AUROC = 0.931; AUPRC = 0.933; balanced accuracy = 0.834), further improved after optimization (balanced accuracy = 0.886; sensitivity = 0.842; specificity = 0.930). Classical CNN architectures demonstrated performance comparable to or exceeding that of more complex hybrid CNN–Transformer models when anatomically focused preprocessing and rigorous optimization were applied. Conclusions: Anatomically constrained preprocessing through breast-mask segmentation substantially enhances deep learning performance and stability in CEM-based breast lesion classification. These findings indicate that input representation quality and training optimization are critical determinants of clinically relevant performance, often outweighing architectural complexity, and may support more reliable AI-assisted decision support in CEM workflows.

View free PDFSource page

Related papers

crossrefBioengineering2026-07-24

Beyond Size and Class Balance: Alpha as a New Dataset Quality Metric for Deep Learning

Josiah D. Couch, Rima Arnaout, Ramy Arnaout

In deep-learning-based image classification, achieving high performance requires diverse training sets. However, the current best practice—maximizing dataset size and class balance—does not guarantee dataset diversity. We hypothesized that, for a given model architecture, perform…

View free PDFSource page
openalexBioengineering2026-07-23

Optimizing Transcutaneous Electrical Stimulation Based on Frequency-Dependent Tissue Modeling and Signal Analysis

Wenzhu Wu, Junquan Tang, Q Wang, Jun Yang

Transcutaneous electrical stimulation (TES) is limited by cutaneous discomfort caused by unavoidable activation of superficial sensory nerves during current delivery to deep targets. While psychophysical studies have empirically identified waveforms that reduce skin sensation, ex…

View free PDFSource page
openalexBioengineering2026-07-23

AI-Driven Robotic PCI: A Perception-to-Action Framework for Coronary Guidewire Control

Zijing Liu, Huanming Xu, Zhendong Liu, Jun Ma, J I N Liu, Pengfei Bi, et al.

Robotic percutaneous coronary intervention (PCI) remains predominantly teleoperated, while the most demanding part of the procedure, guidewire navigation through a moving coronary tree under fluoroscopy, still depends on the continuous human interpretation of vessel anatomy, card…

View free PDFSource page
openalexBioengineering2026-07-23

Quantitative Retinal Vascular Imaging Using Flood-Illuminated Adaptive Optics in Eyes with Lens Opacities

Yash Porwal, Kenaan Elbash, Marlon Diaz, Emily Butler, Saige Oechsli, Laurence Magder, et al.

Flood-illuminated adaptive optics (AO) imaging enables high-resolution quantification of retinal vascular morphology; however, its repeatability in older adults with lens opacities remains poorly characterized. In this prospective study, 40 subjects (66 eyes) underwent flood AO i…

View free PDFSource page
crossrefBioengineering2026-04-15

Optimized Signal Acquisition and Advanced AI for Robust 1D EMG Classification: A Comparative Study of Machine Learning, Deep Learning, and Reinforcement Learning

Anagha Shinde, Virendra Shete, Ninad Mehendale

Electromyography (EMG) signals are critical for prosthetic control, rehabilitation, and human–machine interaction, yet their classification remains challenging due to noise, non-stationarity, and inter-subject variability. This study presents a comprehensive comparative analysis…

View free PDFSource page
crossrefBioengineering2026-04-10

Neuroscience-Inspired Deep Learning Brain–Machine Interface Decoder

Hong-Yun Ou, Takahiro Hasegawa, Osamu Fukayama, Eizo Miyashita

Brain–machine interfaces (BMIs) aim to decode motor intentions from neural activity to enable direct control of external devices. However, most existing decoders rely on monolithic architectures that fail to capture the distinct neural representations of different joint movement…

View free PDFSource page