Machine Vision Intelligence Using Layer-Wise Relevance Backward Propagation For Breast Cancer Diagnosis

Abstract

<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>Machine vision intelligence with layer-wise relevance propagation (LRP) for breast cancer diagnosis is an innovative approach aimed at improving diagnostic accuracy and efficiency. Traditional methods for diagnosing breast cancer often rely on the manual interpretation of medical images, which can be time-consuming and prone to human error. This study proposes the development of a machine vision model utilising LRP designed to analyse mammographic images for the diagnosis of breast cancer. Various machine learning models, including Convolutional Neural Networks (CNN), VGG19, VGG16, ResNet50, EfficientNet, and Vision Transformer, were trained on a breast cancer ultrasound dataset of mammograms. Features such as texture, density, and shape were extracted from each image for analysis. The different models achieved high accuracies of $\mathbf{9 9 \%}$ $\mathbf{9 1 \%}, \mathbf{8 0 \%}, \mathbf{9 9 \%}, \mathbf{9 5 \%}$, and $\mathbf{8 9 \%}$, respectively. LRP is then used to explain the model’s classification by highlighting image regions most relevant to its decision. This approach aims to attain high diagnosis accuracy while also offering interpretable insights into the model’s logic, potentially increasing its reliability in clinica contexts. By automating the interpretation of mammographic images, this has the potential to improve diagnosis accuracy and patient outcomes.<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>Makerere University Research and Innovations Fund, under Digital Domiciliary Platform to Standardize Data Capture in Community-based Midwifery Practices and Quality of Care (DC4Midwifery)