In the human healthcare area, leukocytes are very important blood cells for the diagnosis of different pathologies, like leukemia. Recent technology and image-processing methods have contributed to the image classification of leukocytes. Especially, machine learning paradigms have been used for the classification of leukocyte images. However, reported models do not leverage the knowledge produced by the classification of leukocytes to solve similar tasks. For example, the knowledge can be reused to classify images collected with different types of microscopes and image-processing techniques. Therefore, we propose a meta-learning methodology for the classification of leukocyte images using different color constancy methods involving previous knowledge. Our methodology is trained with a specific task at the meta-level, and the knowledge produced is used to solve a different task at the base-level. For the meta-level, we implemented meta-models based on Xception, and for the base-level, we used support vector machine classifiers. Besides, we analyzed the Shades of Gray color constancy method commonly used in skin lesion diagnosis and now implemented for leukocyte images. Our methodology, at the meta-level, achieved 89.28% for precision, 95.65% for sensitivity, 91.78% for F1-score, and 94.40% for accuracy. These scores are competitive regarding the reported state-of-the-art models, especially the sensitivity which is very important for imbalanced datasets, and our meta-model outperforms previous works by +2.25%. Additionally, for the basophil images that were acquired from a chronic myeloid leukemia-positive sample, our meta-model obtained 100% for sensitivity. Moreover, we present an algorithm that generates a new conditioned output at the base-level obtaining highly competitive scores of 91.56% for sensitivity and F1 scores, 95.61% for precision, and 96.47% for accuracy. The findings indicate that our proposed meta-learning methodology can be applied to other medical image classification tasks and achieve high performances by reusing knowledge and reducing the training time for new similar tasks.
Read full abstract