فهرست نشریات علمی دسترسی آزاد ایران

This research pioneers an innovative and genuinely promising methodology to tackle the pervasive challenge of overfitting in the domain of breast cancer tumor analysis, a common obstacle, particularly when confronted with limited datasets. The central aim is to significantly enhance diagnostic accuracy and fortify overall model robustness. This is achieved through a synergistic integration of transfer learning principles with the sophisticated data augmentation capabilities afforded by Deep Convolutional Generative Adversarial Networks (DCGANs). The fundamental problem is the scarcity of training samples, a frequent predicament in medical imaging. This scarcity can lead to models that excel on familiar training data but falter significantly when encountering new, unseen patient cases, undermining their clinical utility. To surmount this, the researchers judiciously utilized the standard MIAS (Mammographic Image Analysis Society) database. A DCGAN architecture, known for its proficiency in generating realistic images by pitting a generator network against a discriminator network, was employed. This network learns the underlying patterns and distributions of the original data to produce synthetic mammographic images. This process effectively expanded the training pool, resulting in the creation of 10,000 high-quality synthetic data points. Crucially, these synthetic images were designed to realistically mimic the complex and often subtle characteristics inherent in actual breast tumor images found within the MIAS dataset, ensuring they contribute meaningfully to model training. These newly generated synthetic samples, combined with the limited original MIAS data, formed an augmented dataset. This enriched dataset was then used to train the YOLOV11m neural network architecture. The application of transfer learning was pivotal here. This technique allowed the YOLOV11m model to benefit from knowledge pre-acquired from training on larger, more general datasets, significantly enhancing its learning efficiency and overall performance. This is especially critical when the original domain-specific dataset (like MIAS) is small, as it provides a robust foundational understanding of visual features. The experimental results compellingly demonstrated the remarkable efficacy of this integrated methodology. The YOLOV11m model, when trained on this augmented dataset, achieved an impressive 99.1% accuracy in distinguishing between benign and malignant tumors.