MiniExtractor-PHyMiPS: A Parameter-Efficient Pipeline With Compact Feature Extractor, MI, and PCA for Breast Cancer Detection

Breast ultrasound (BUS) is an essential imaging technique for the early detection of breast cancer, particularly in women with dense breast tissue. However, the high interclass similarity and variability in tumor morphology pose significant challenges for automated classification. To address this, MiniExtractor-PHyMiPS, a hybrid pipeline that integrates deep learning for complex feature extraction and conventional machine learning for robust classification, is proposed. This hybrid pipeline consists of a compact convolutional neural network (CNN)-based MiniExtractor to identify complex features from BUS images. Subsequently, mutual information (MI) analysis is used to compute feature relevance scores, and the top-ranked features are selected. These selected features are then transformed using principal component analysis (PCA) to reduce the dimensionality of the feature space before classification. A support vector machine (SVM) classifier is used for the final decision-making. Extensive experiments were conducted on two publicly available datasets: the Breast Ultrasound Images (BUSI) dataset and the Ultrasound Breast Images for Breast Cancer (UBIBC) dataset. The proposed pipeline achieved 96.90% accuracy on BUSI and 95.81% on UBIBC, surpassing several state-of-the-art deep learning models, including ResNet50, DenseNet201, and XceptionNet, while using only 448 parameters in the MiniExtractor feature extractor. Furthermore, the proposed hybrid pipeline demonstrated consistent bidirectional cross-dataset generalization between BUSI and UBIBC, while maintaining a compact end-to-end deployment footprint. Although the MiniExtractor itself contains only 448 parameters, the complete deployed pipeline—including PCA transformation (1870 parameters) and the RBF-SVM classifier (17,256 parameters with 493 support vectors)—results in a total of approximately 20k parameters (19,574). In addition, the end-to-end inference complexity is approximately 0.884 MFLOPs per sample, with an average CPU inference time of 0.0478 ms per image, which is substantially lower than conventional end-to-end deep learning models. The findings of this study reveal the high efficacy and robustness of the MiniExtractor-PHyMiPS in achieving superior performance of automated breast cancer detection via ultrasound images, making it well-suited for real-world clinical applications, including deployment in low-resource settings and mobile diagnostic platforms.