Auto Machine Learning Based on Genetic Programming for Medical Image Classification

Medical imaging classification is an area that has taken relevance in recent years due to the capability to support the medical specialist at the time of diagnosis. However, there are different instruments to obtain images from the body, and each body organ is captured differently due to its chemical composition. In this way, there are some difficulties in working with different imaging modalities. Firstly, using different functions or methods to extract features from the images is necessary. Secondly, the classification performance depends on the relevant features extracted from the images, and thirdly, it is necessary to find the classifier that performs with the minimum error. Following the concept of Auto-Machine Learning (AutoML), where the feature engineering and the hyperparameter tuning of the classifier are done automatically, this work proposes an automated approach for feature extraction and image classification based on Genetic Programming. The approach modifies the functions and their parameters and the hyperparameters for the classifier. The results show that the approach can deal with different imaging modalities, demonstrating that feature extraction is necessary to increase the classification performance. For X-ray images, it achieves a classification accuracy of 0.99, and for computerized tomography, it achieves an accuracy of 0.96. On the other hand, the solutions given by the approach are easily reproducible and easy to interpret.