Transfer Learning for Sickle Cell Anemia and Trait Classification

Sickle cell disease is one of the most prevalent inherited blood disorders. The majority of the population suffering from this disorder are the active carrier of the disease (sickle cell trait) and are unaware of their health status. To have effective prevention of the spread of disease proper demarcation between disease and trait is required. The existing pathological methods for disease diagnosis are costly and time-consuming while most of the machine learning-based method focuses on normal versus abnormal cell classification. In this study transfer learning of pre-trained AlexNet model is proposed for classification of disease versus trait cases, a very first approach towards the sickle cell diseases subtype classification with the aid of machine learning and image processing tools. Also, the performance of the model is evaluated under various data division protocols, hold-out, 5-fold, 10-fold respectively. The study is conducted on a newly prepared database of 67 traits and 23 disease cases. The proposed system shows the highest classification accuracy of 95.5% with 10-fold data division protocol. Other performance parameters used for evaluation are precision, sensitivity, specificity, neg predicted value and ROC curve. In addition, the study examines a practical feature of the system by assessing it with fewer training samples. Also, the findings of the study suggest that transfer learning appears to be a helpful strategy when the availability of a medical dataset is restricted.