Journal of Advanced Robotics, Autonomous Systems and Human-Machine Interaction
Detection of Abnormalities in Blood Cells Using a Region-Based Segmentation Approach and Supervised Machine Learning Al-gorithm
Abstract
Nagueu Djambong Lionel Perin, Waku Kouomou Jules, Hippolyte Kenfack Tapamo and Jimbo H Claver
Screening (slide reading stage) is a manual human activity in cytology which consists of the inspection or analysis by the cytotechnician of all the cells present on a slide. Segmentation of blood cells is an important research question in hematology and other related fields. Since this activity is human-based, detection of abnormal cells becomes difficult. Nowadays, medical image processing has recently become a very important discipline for computer-aided diagnosis, in which many methods are applied to solve real problems. Our research work is in the field of computer-assisted diagnosis on blood images for the detection of abnormal cells. To this end, we propose a hybrid segmentation method to extract the correct shape from the nuclei to extract features and classify them using SVM and KNN binary classifiers. In order to evaluate the performance of hybrid segmentation and the choice of the classification model, we carried out a comparative study between our hybrid segmentation method followed by our SVM classification model and a segmentation method based on global thresholding followed by a KNN classification model. After this study, it appears from the experiments carried out on the 62 images of blood smears, that the SVM binary classification model gives us an accuracy of 97% for the hybrid segmentation and 57% in the global thresholding and 95 % for the KNN Classification Model. As our dataset was not balanced, we evaluated precision, recall, F1 score and cross validation with the Stratified K-Fold cross validation algorithm of each of these segmentation methods and classification models. We obtain respectively: 93.75%; 98.712% and 99% for hybrid segmentation reflecting its effectiveness compared to global fixed threshold segmentation and KNN classification model. To evaluate the performance of these models we obtained the following results: 77% of mean accuracy in the SVM and 61% of mean accuracy in the KNN, 84% of mean test accuracy in the SVM and 74% mean test accuracy in the KNN making the best performing SVM model