Heuristic optimization of preprocessing and model parameters to improve classification success in medical images

Medical imaging techniques are used to detect, monitor and treat diseases. However, experts who make diagnoses using medical images are also affected by human factors such as forgetfulness and fatigue. Models developed with deep learning technology, a sub-branch of artificial intelligence, are used in the diagnosis and classification of many diseases. In this study, the most appropriate models and parameters were determined by heuristic optimization by using preprocessing techniques to increase the classification success of medical images. For the classification of breast cancer datasets (MKVS-1, MKVS-2), lung nodule dataset (ANVS), lung cancer dataset (AKVS) and diabetic retinopathy (DR) eye disease, datasets that have a significant place in the literature were used. A contrast-limited adaptive histogram equalization (CLAHE) preprocessing technique was first applied to the data. Various filters were then applied to these data sets, which were divided into original and CLAHE applied data. After the data was prepared, an optimization study was performed on ANVS with grid search (IA), differential evolution (DE) and particle swarm optimization (PSO) methods. After obtaining the highest accuracy on ANVS with the IA method, DE and PSO algorithms were run 10 times and compared. After DE optimization was found to be more successful, five data sets were run 10 times with the DE algorithm. The highest accuracy scores were obtained in its form without data augmentation in the ANVS, MKVS-1 and DR datasets, and running 10 times with the DE algorithm after in its form applying data augmentation in the AKVS and MKVS-2 datasets. Accuracy of 0,9903 for ANVS, 0,813 for MKVS-1, 0,9141 for DR data set, 0,95 for AKVS and 0,8807 for MKVS-2 was obtained. In comparison with other studies using the same data sets, it was observed that more successful results were obtained for ANVS and AKVS compared to the literature. In terms of accuracy, MKVS-1 and DR datasets obtained results that are very close to the results in the literature, and in other classification metrics, more successful results were obtained compared to other studies.