Deep Learning Approaches For Analyzing Microscopic Peripheral Blood Cell Images.

Abstract

The diagnosis of numerous hematological disorders depends on the precise and effective classification of peripheral blood cells. However, in high-throughput or resource-constrained contexts, the manual evaluation of peripheral blood smear (PBS) pictures is laborious, prone to errors, and requires expert interpretation. Using a dataset of 17,092 labeled photos, this work explores deep learning techniques to automatically classify eight blood cell types: basophil, eosinophil, erythroblast, immature granulocytes, lymphocyte, monocyte, neutrophil, and platelet. This dataset was used to refine and assess advanced pre-trained models, such as MobileNetV2, ResNet50, ResNet101V2, InceptionV3, VGG16, and EfficientNetB3. MobileNetV2 and EfficientNetB3 were combined to create a hybrid model, which was then improved and assessed. Out of all the models, the Hybrid model had the highest overall accuracy of 95% and the best F1 score of 0.95 on the macro level, and the best loss of 0.15. Most symbolically, the performance of the platelet and eosinophil classes was virtually perfect with the F1-scores of 1.00 and 0.97, accordingly. Even if MobileNetV2 and EfficientNetB3 had a very high efficiency and accuracy, minor deficiencies were revealed for example in immature granulocytes and monocytes classifications; however, Overall all models proposed a fairly high recall. To enhance the results of model generalization and enhance the quality of images, strategies like data augmentation, noise elimination, and image normalization were employed. This research work provides the much-needed confirmation of MobileNetV2 as a very good substitute while providing empirical evidence of the Hybrid Model’s ability to scale higher than models trained on a single source in terms of accuracy and robustness. For future works, more efforts will be devoted to creating a larger dataset, refining existing hybrid methods, and enhancing the recall rate of the inferior classification. By enhancing the technique of automatic classification of peripheral blood cells, this work has the potential of developing better and more stylistic clinical and diagnostic arrangements.