Classification of chest ultrasound images using artificial i̇ntelligence techniques

In this study, a classification system was developed to distinguish between benign and malignant lesions in breast cancer diagnosis using ultrasound images through various pretrained convolutional neural network (CNN) models. The research primarily utilized a total of 9,016 ultrasound images obtained from the Ultrasound Breast Images for Breast Cancer (UBIBC) dataset. The dataset consists of a balanced distribution of 4,074 benign and 4,042 malignant training images, along with 500 benign and 400 malignant test images. The original resolution of the images is 700×460 pixels, and all images were resized to 224×224×3 and normalized to be compatible with the CNN architectures. The RGB images with 8-bit depth per channel underwent preprocessing prior to model training. In the modeling phase, multiple CNN architectures such as ResNet50, DenseNet121, EfficientNetB0, MobileNetV3, and ShuffleNetV2 were employed. Feature extraction was performed using transfer learning, and the extracted features were classified using various supervised machine learning algorithms, including KNN, Cubic SVM, and Linear Discriminant. All experiments were conducted using the Python programming language on Google Colab with T4 GPU support. The models were evaluated based on performance metrics including accuracy, precision, recall, F1-score, training time, and prediction speed. According to the experimental results, the EfficientNetB0 + Cubic SVM architecture achieved a high accuracy of 96.73%, while the ResNet50 + KNN (Fine) model also stood out with 96.18% accuracy. Furthermore, features extracted from the combined lightweight architectures MobileNetV3 + ShuffleNetV2 reached 98.12% accuracy when classified using KNN Fine, producing the best results in the study. This combination was notable not only for its high classification success but also for its minimal training time (0.01 seconds). Upon examination of ROC curves, all models had AUC values above 0.95, and false negative rates were found to be remarkably low, which is particularly critical in medical applications and diagnostic safety. In conclusion, feature extraction from different CNN architectures followed by classification using classical machine learning algorithms yields highly successful results in breast cancer diagnosis based on ultrasound images. In particular, the MobileNetV3 + ShuffleNetV2 + KNN Fine combination provides a lightweight, fast, and high-accuracy solution suitable for mobile and low-resource medical systems. This study demonstrates that deep learning-based diagnostic systems can offer more accurate, faster, and more applicable alternatives compared to traditional methods. For future work, the model's generalizability will be explored by applying it to larger and more diverse datasets. KEYWORDS: CNN, Deep Learning, Breast Cancer, Classifiers, Artificial Intelligence