DISTRIBUTED FEATURE EXTRACTION WITH AUTOENCODERS FOR BREAST CANCER DETECTION

Abstract

In this study, a distributed feature extraction pipeline was developed for breast cancer detection us-ing autoencoders. Mammogram images were first derived from the RSNA dataset by converting DICOM files into PNG format. Twelve convolutional autoencoders were trained, where all layers were convolutional except for the bottleneck layer, which was defined as a fully connected layer. This layer served as the compressed feature vector. Variations across models were introduced by modifying the number of convolutional layers in encoders/decoders (3, 4, or 5) and the dimension-ality of the feature vector (128, 256, 512, or 1024). Training was conducted using mean squared error loss in a synchronous multi-worker setup on a four-node CPU cluster utilizing the Keras framework. After training, the extracted features were evaluated using three classification models: logistic regression, XGBoost, and CatBoost. The performance of each feature vector configuration was assessed based on accuracy, precision, recall, and F1-score. Through comparative analysis, the effectiveness of different vector sizes and model complexities in representing diagnostic features was determined. This approach demonstrated the feasibility of scalable, distributed feature extrac-tion for high-resolution medical imaging tasks, offering a practical framework for future breast cancer detection systems.