Breast cancer is one of the most common non-communicable diseases affecting women worldwide and remains the leading cause of cancer-related deaths among women. According to international data, more than one-third of all cancer cases in women originate from breast cancer. Although medical technology has advanced significantly, early detection remains a crucial factor in determining the success of treatment. More than 90% of cases can be cured if diagnosed at an early stage. However, in clinical practice, diagnosis is not always straightforward, as imaging modalities such as ultrasonography (USG), mammography, and magnetic resonance imaging (MRI) require detailed interpretation by highly trained medical professionals. Differences in texture, size, and shape of lesions often cause ambiguity, potentially leading to misdiagnosis.
The rapid advancement of artificial intelligence (AI), particularly in deep learning, has opened new opportunities to assist in the automatic diagnosis of breast cancer. Various models, such as Convolutional Neural Networks (CNNs) and Vision Transformers, have been applied in medical image classification. However, these models still face challenges in identifying small lesions and handling low-contrast images. To overcome these limitations, this study introduces BreastDiff, a multi-condition guided diffusion model developed to enhance the system’s ability to recognize and classify different types of breast cancer lesions across various imaging modalities.
The effectiveness of using deep learning technology
This study proposes a novel model called BreastDiff, developed to improve the accuracy of breast cancer classification across multiple imaging types, including ultrasonography (USG), mammography, and MRI. The model adopts a diffusion-based deep learning approach — a technique that progressively adds and removes noise during training to learn complex data distributions.
The main objective of this method is to enable the model to recognize not only global patterns within an entire image but also focus on relevant local areas, such as the lesion or suspicious mass. This dual-level focus allows the system to perform classification more accurately, even for small lesions or images with low contrast.
Three Complementary Mechanisms
In principle, BreastDiff is built upon three complementary mechanisms. The first is Multi- Condition Guided Diffusion (Diff-MCG), which integrates global information from the entire breast image with local details from the lesion area. This approach strengthens essential features without losing overall anatomical context.
The second is UC-Net with Tri-Channels Attention, which employs a three- directional attention mechanism involving spatial attention, channel attention, and coordinate attention. These mechanisms work together to enhance the model’s sensitivity to fine details in the lesion area, enabling the system to distinguish between healthy tissue and cancerous regions even when the visual differences are subtle.
The third component is the Contrastive Maximum Mean Discrepancy (MMDC) Loss Function, a novel loss function that combines contrastive loss and maximum mean discrepancy. This approach allows the model to distinguish between image classes more effectively while improving its generalization ability, ensuring stability when dealing with diverse data variations.
The combination of these three mechanisms makes BreastDiff both efficient and highly accurate in identifying breast cancer patterns, particularly in low-quality or noisy images commonly found in clinical datasets. Through this approach, the system not only builds upon existing datasets explored in previous studies but also develops a more contextual and adaptive understanding of diverse medical image characteristics. Consequently, BreastDiff demonstrates strong potential as a reliable and applicable model to support automatic breast cancer diagnosis in modern clinical environments.
Methods and Results
This study utilized three types of breast imaging datasets: BUSI (ultrasound), CMMD (mammography), and BreastDM (MRI). The BUSI dataset consisted of 437 benign and 210 malignant images, the CMMD dataset contained 5,210 benign and 2,619 malignant images, and the BreastDM dataset included 2,448 benign and 2,505 malignant images. All datasets were divided into training, validation, and testing sets with proportions of 80%, 10%, and 10%, respectively.
Model training was conducted in an Ubuntu 20.04 environment using an NVIDIA Tesla T4 GPU (16 GB) with optimization via stochastic gradient descent (SGD) for 100 epochs. Model performance was evaluated using Accuracy, Precision, Recall, and F1-Score metrics.
The results showed that BreastDiff achieved an accuracy of 96.16% on the BUSI dataset and 96.89% on the CMMD dataset, representing an approximately 12.9% improvement over baseline models and an average of 2.6% higher performance than other state-of-the-art models, including Swin-Transformer, D-Cube, and DIFFMIC.
Furthermore, efficiency tests demonstrated that the quantization and pruning processes reduced the model size from 613 MB to 324 MB, with only a 3.9% reduction in accuracy. This finding indicates that BreastDiff remains stable even after compression, making it more efficient and practical for implementation on portable medical devices in clinical settings.
Author: Yuxin You, Chenyi Zhuang, Hong-Seng Gan, Riries Rulaningtyas, Muhammad Hanif Ramlee, dan Asnida Abdul Wahab
Detailed information from this research can be seen in our writing at:
You, Y., Zhuang, C., Gan, H.-S., Rulaningtyas, R., Ramlee, M. H., & Wahab, A. A. (2025). BreastDiff: A multi-condition guided diffusion model for breast cancer classification in diverse modalities. Biomedical Signal Processing and Control, 112, 108708. https://doi.org/10.1016/j.bspc.2025.108708
