Advancing Breast Cancer Detection: A Comparison of PCA and LDA Methods in Analyzing Ultrasound Imagery

Abstract

Early and accurate detection of breast cancer via ultrasound imaging is essential, yet the high dimensionality of raw ultrasound features can hinder classifier performance and increase computational burden. Comparison between Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) for feature reduction in a breast-cancer ultrasound diagnostic pipeline, alongside t-SNE for exploratory visualization. The research utilized 1,200 breast ultrasound images with 400 benign, 400 malignant, and 400 normal images obtained from Baheya Hospital (Cairo, Egypt). Minority classes were balanced using data augmentation techniques like rotation and flipping. PCA reduced the data to 172 components, preserving 90% of data variance, while LDA used two components. t-SNE generated a two-dimensional visual representation. Classifiers, including Support Vector Machine (SVM), Random Forest (RF), and XGBoost, were trained on: (a) the full feature set, (b) PCA-reduced data, and (c) LDA-reduced data. Evaluation metrics included precision, recall, and F1-score. Compression ratio and signal-to-noise ratio (SNR) measured image compression via PCA.Without reduction, XGBoost achieved the highest F1-score (76.97%), precision (77.40%), and recall (76.55%). PCA yielded a modest precision gain (XGBoost: 78.65%) but reduced recall and net F1-score (76.37%). LDA significantly degraded performance (XGBoost F1: 63.99%; RF F1: 60.13%; SVM F1: 42.05%). PCA compression reduced image size by 2.68x with an SNR of 48.91 dB, while LDA offered no compression benefit. t-SNE visualization revealed clear non-linear class clusters, underscoring the dataset’s intrinsic complexity. For ultrasound-based breast cancer diagnosis, preserving full high-dimensional features and using a powerful non-linear model (e.g., XGBoost) yields optimal accuracy. PCA is best reserved for storage or runtime efficiency, LDA for scenarios with very low dimensional constraints, and t-SNE for exploratory data analysis. This comparative study highlights that dimensionality reduction may harm performance in complex imaging data and recommends context-specific use of PCA and LDA to avoid loss of critical diagnostic information.