Multi-task interaction learning for accurate segmentation and classification of breast tumors in ultrasound images

Physics in Medicine and Biology, Journal Year: 2025, Volume and Issue: 70(6), P. 065006 - 065006

Published: Jan. 24, 2025

Objective.In breast diagnostic imaging, the morphological variability of tumors and inherent ambiguity ultrasound images pose significant challenges. Moreover, multi-task computer-aided diagnosis systems in imaging may overlook relationships between pixel-wise segmentation categorical classification tasks.Approach.In this paper, we propose a learning network with deep inter-task interactions that exploits inherently relations two tasks. First, fuse self-task attention cross-task mechanisms to explore types interaction information, location semantic, In addition, feature aggregation block is developed based on channel mechanism, which reduces semantic differences decoder encoder. To exploit further, our uses an circle training strategy refine heterogeneous help maps obtained from previous training.Main results.The experimental results show method achieved excellent performance BUSI BUS-B datasets, DSCs 81.95% 86.41% for tasks, F1 scores 82.13% 69.01% respectively.Significance.The proposed not only enhances all tasks related tumor but also promotes research learning, providing further insights clinical applications.

Language: Английский

---

Empowering early detection: artificial intelligence as a tool for breast cancer diagnosis

Elsevier eBooks, Journal Year: 2025, Volume and Issue: unknown, P. 121 - 145

Published: Jan. 1, 2025

Language: Английский

---

NMTNet: A Multi-task Deep Learning Network for Joint Segmentation and Classification of Breast Tumors

Deleted Journal, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 19, 2025

Segmentation and classification of breast tumors are two critical tasks since they provide significant information for computer-aided cancer diagnosis. Combining these leverages their intrinsic relevance to enhance performance, but the variability complexity tumor characteristics remain challenging. We propose a novel multi-task deep learning network (NMTNet) joint segmentation tumors, which is based on convolutional neural (CNN) U-shaped architecture. It mainly comprises shared encoder, multi-scale fusion channel refinement (MFCR) module, branch, branch. First, ResNet18 used as backbone in encoding part feature representation capability. Then, MFCR module introduced enrich depth diversity. Besides, branch combines lesion region enhancement (LRE) between encoder decoder parts, aiming capture more detailed texture edge irregular improve accuracy. The incorporates fine-grained classifier that reuses valuable discriminate benign malignant tumors. proposed NMTNet evaluated both ultrasound magnetic resonance imaging datasets. achieves dice scores 90.30% 91.50%, Jaccard indices 84.70% 88.10% each dataset, respectively. And accuracy 87.50% 99.64% corresponding datasets, Experimental results demonstrate superiority over state-of-the-art methods tasks.

Language: Английский

---

Variational mode directed deep learning framework for breast lesion classification using ultrasound imaging

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: April 24, 2025

Breast cancer is the most prevalent and second cause of related death among women in United States. Accurate early detection breast can reduce number mortalities. Recent works explore deep learning techniques with ultrasound for detecting malignant lesions. However, lack explanatory features, need segmentation, high computational complexity limit their applicability this detection. Therefore, we propose a novel ultrasound-based lesion classification framework that utilizes two-dimensional variational mode decomposition (2D-VMD) which provides self-explanatory features guiding convolutional neural network (CNN) mixed pooling attention mechanisms. The visual inspection these demonstrates explainability terms discriminative lesion-specific boundary texture decomposed modes benign images, further guide enhanced classification. proposed classify lesions accuracies 98% 93% two public datasets 89% an in-house dataset without having to segment unlike existing techniques, along optimal trade-off between sensitivity specificity. 2D-VMD improves areas under receiver operating characteristics precision-recall curves by 5% 10% respectively. method achieves relative improvement 14.47%(8.42%) (mean (SD)) accuracy over state-of-the-art methods one dataset, 5.75%(4.52%) another comparable performance methods. Further, it computationally efficient reduction [Formula: see text] floating point operations as compared

Language: Английский

---

A multi-task framework for breast cancer segmentation and classification in ultrasound imaging

Computer Methods and Programs in Biomedicine, Journal Year: 2024, Volume and Issue: 260, P. 108540 - 108540

Published: Dec. 4, 2024

Ultrasound (US) is a medical imaging modality that plays crucial role in the early detection of breast cancer. The emergence numerous deep learning systems has offered promising avenues for segmentation and classification cancer tumors US images. However, challenges such as absence data standardization, exclusion non-tumor images during training, narrow view single-task methodologies have hindered practical applicability these systems, often resulting biased outcomes. This study aims to explore potential multi-task enhancing lesions.

Language: Английский

---

Exploiting K-Space in Magnetic Resonance Imaging Diagnosis: Dual-Path Attention Fusion for K-Space Global and Image Local Features

Bioengineering, Journal Year: 2024, Volume and Issue: 11(10), P. 958 - 958

Published: Sept. 25, 2024

Magnetic resonance imaging (MRI) diagnosis, enhanced by deep learning methods, plays a crucial role in medical image processing, facilitating precise clinical diagnosis and optimal treatment planning. Current methodologies predominantly focus on feature extraction from the domain, which often results loss of global features during down-sampling processes. However, unique representational capacity MRI K-space is overlooked. In this paper, we present novel K-space-based dual-path attention fusion network. Our proposed method extracts data fuses them with local domain using mechanism, thereby achieving accurate segmentation for diagnosis. Specifically, our consists four main components: an image-domain module, decoder. We conducted ablation studies comprehensive comparisons Brain Tumor Segmentation (BraTS) dataset to validate effectiveness each module. The demonstrate that exhibits superior performance diagnostics, outperforming state-of-the-art methods improvements up 63.82% HD95 distance evaluation metric. Furthermore, performed generalization testing complexity analysis Automated Cardiac Diagnosis Challenge (ACDC) cardiac dataset. findings indicate robust across different datasets, highlighting strong generalizability favorable algorithmic complexity. Collectively, these suggest holds significant potential practical applications.

Language: Английский

---