Classification of non-Hodgkin lymphomas based on sample entropy signatures

Computational systems to provide studies and diagnoses of non-Hodgkin’s lymphomas have been increasingly developed to assist specialists in their decision-making. On the other hand, the approaches have not yet fully explored the sample entropy to assess the disease’s characteristics, with indications of the best features and techniques to distinguish the main groups of lymphomas. Herein, we present a method that considers sample entropy signatures to study chronic lymphocytic leukemia, follicular lymphoma, and mantle cell lymphoma, involving images after applying multiple segmentation techniques and color normalization models. Texture signatures were defined as feature curves obtained from multiple observations, by associating different parameters. Each signature was calculated for windows of size m and ranged from 1 to 4 (0.06–0.40) tolerance levels (r). The behavior of each signature was determined from the area under the curve, skewness, maximum entropy value and area ratio, representing the signature metrics. This approach aimed to improve the quantitative ability of this descriptor, with new interpretations and associations. Thus, all features were provided for different functions, such as lazy learning, trees, genetic evolution, particle swarm and animal behavior-based approaches. The best association was tested after adding noise levels of 20%, 40% and 80%. The performances achieved with the proposal were accuracy rates between 98.72% and 99.60%. We were able to obtain an accuracy of 98.84% for multi-class classification using only ten features, which is an amount considerably lower than most of the state-of-the-art approaches. The proposed approach presents relevant advances for the investigation of computer-aided diagnosis systems since it provides accurate results and does not require preprocessing steps. Moreover, signature metrics in association with evolutionary approaches, which provided the best results, are among the main contributions of this paper, since the details of the most relevant features with a full understanding of the solutions are useful for the study and pattern recognition of non-Hodgkin’s lymphomas.