Evrişimsel sinir ağları kullanılarak meme kanseri tespiti

Breast cancer continues to be one of the most popular as well as life-threatening malignancies because it affects women globally, demanding timely along with accurate diagnostic interventions for improved survival rates and therapeutic outcomes. Ultrasound imaging acts as a critical diagnostic tool because of its non-invasiveness as well as cost-effectiveness, plus efficacy when visualizing dense breast tissues, especially in low-resource settings where mammography and MRI may not be readily available. Ultrasound images get diagnostic interpretation that is customarily variable, with limitations in sensitivity and specificity. These issues do underscore the need for such smart automated decision-support systems. For breast ultrasound image classification, we propose a novel fully customized Convolutional Neural Network (CNN) architecture carefully optimized. This study's classification includes three clinically important categories: normal, malignant, and benign. We used the Breast Ultrasound Images (BUSI) dataset, and we tackled intrinsic issues like small dataset size, class imbalance, and image noise via a diverse augmentation pipeline, which included rotation, flipping, translation, shear transformation, and brightness adjustment. Because the model could generalize effectively across variations of the real world, the resulting dataset of 6,000 images greatly improved the diversity and representativeness of training data. The proposed CNN architecture integrates state-of-the-art deep learning techniques that include hierarchical feature extraction, batch normalization, dropout regularization, along with adaptive learning rates through a ReduceLROnPlateau scheduler. The best performing model did achieve 98.83% macro-averaged F1-score and also test accuracy. This result was thanks to wide-ranging hyperparameter tuning across learning rates, batch sizes, and input resolutions. Although trained without pre-existing knowledge, the model showed better 6 performance when evaluated against popular transfer learning models like MobileNetV2, VGG16, VGG19, InceptionV3, ResNet50, EfficientNetB0, DenseNet121, and Xception. Computational efficiency, interpretability, as well as high diagnostic accuracy can be achievable via domain-specific, lightweight CNN models according to these results. According to our findings, tailored deep learning solutions play a critical role, highlighting their transformative potential in early breast cancer detection and in helping medical imaging, especially inside resource-constrained environments.