SUMMARY
The model development for breast thermal image classification can be done using deep learning methods, especially the convolutional neural network (CNN) architecture. This article focuses on adapting a trained CNN (trained model) on a mobile application for binary classification of breast thermal images into normal and abnormal classes. The CNN model applied in this study was based on ShuffleNet, called BreaCNet, with a learning weight of 1028 filters generated from training on images downloaded from the Database for Mastology Research (DMR) and a model size of 22 MB. The model must be converted into a mobile application to enable a trained model to be adapted into a mobile platform. The BreaCNet model was built using MatLab; thus, the stages in the adaptation process consisted of converting the model into ONNX file format, converting ONNX files into Tensorflow files, and Tensorflow files into Tensorflow Lite format. However, not all nodes are fully supported by MATLAB. The shuffle node on ShuffleNet cannot be fully exported using ExportToOnnx, so it needs to be re-defined with a placeholder named “MATLAB PLACEHOLDER”. In addition to the model conversion process, this article describes the user interaction process with the application using UML diagrams and application feature menu designs. The application was also tested on 20 thermal images of the breast. The testing results show that the application can perform the image classification process on mobile devices in less than 1 second with an accuracy rate of 85%. Finally, the breast thermal image screening application has been successfully built by directly interpreting the thermal image of the breast on a mobile device to keep the user data private.