Analysis of Convolutional Autoencoder and U-Net forBreast Tumor Detection in Mammogram Images

Abstract

Breast cancer is one of the most frequently diagnosed cancers among women worldwide. Early detection through mammogram analysis is important, but manual interpretation can be challenging because of low image contrast and variability in radiological assessment. This study evaluates the implementation of Convolutional Autoencoder (CAE) and U-Net architectures for mammogram image analysis and integrates both models into a Streamlit-based web application. The CBIS-DDSM dataset was used, comprising 723 training images and 181 testing images. The research stages included preprocessing, model training, evaluation, result analysis, and web application implementation. The U-Net model achieved a Dice Score of 0.4186, an Intersection over Union (IoU) of 0.2647, a precision of 0.5254, a recall of 0.3479, an F1-score of 0.4186, and an accuracy of 0.8955. Meanwhile, the CAE model employed a reconstruction-error threshold for anomaly detection, producing a mean anomaly score of 1.4779% and identifying 31 anomalous images among the 181 test images, using a reported threshold of 1.7578%. These findings indicate that U-Net is suitable for supervised lesion segmentation and localization, whereas CAE provides an annotation-independent approach to reconstruction-based anomaly analysis. However, further evaluation is required to establish the reliability of anomaly detection and the clinical applicability of both models.