AECNN: A Dual-Stage Deep Learning Framework for Accurate and Efficient Ultrasound-Based Breast Cancer Detection

Ultrasound-based breast cancer detection presents unique challenges due to the inherent variability in tissue texture, image quality, and tumor boundary definitions. We introduce a new dual-stage deep learning model with autoencoder and Convolutional Neural Network, termed AECNN (AutoEncoder-Convolutional Neural Network), and investigate the potentials of the newly presented architecture in terms of accurate and consistent breast cancer diagnosis on grayscale ultrasound images. The framework utilizes a compressed latent representation achieved by a high-dimensional input through an unsupervised autoencoder into compact, informative latent information to the supervised 1D convolutional neural network to perform classification. This decoupling of feature extraction and classification can be used to enable the model to remove meaningless noise and concentrate on clinically important patterns. The ultrasound dataset used in this study contains images labeled as benign, malignant, and normal, collected from a publicly available Kaggle repository. Generalization was increased by performing such preprocessing as normalization, grayscale enhancement, and augmentation. After the classification, the AECNN model gained an outstanding result of 98.97, which exceeded several classic CNNs and pre-trained networks, such as VGG16, ResNet50, and DenseNet121. The model showed very high precision (99.10), recall (98.80), and F1-score (98.95) besides the accuracy. It is also advantageous of the framework since the inference is faster than in a deeper network and this allows deployment in real time to clinical decision-support systems. These findings reveal that the AECNN model is very effective to separate between malignant cases and benign and normal cases which makes it a probable source of contention in the use of ultrasound in the diagnosis of breast cancer.