Classification of Erythrocyte Aggregation in Sparse B-mode Ultrasound Image Samples via Stacked Ensemble Deep Learning Models

In this paper, we propose a stacked ensemble heterogeneous deep learning model that combines the prediction results of multiple top-performing single base learners to classify erythrocyte aggregation based on B-mode images of varying sizes under conditions of limited sample size. First, the collected original B-mode images are cropped to create large-size and small-size image datasets. Second, the Swin-transformer, DenseNet121, ConvNext, ResNet34, MobileViT, Efficient, and EfficientNetV2 are individually trained on two distinct datasets, following which the models are evaluated and ranked on their classification metrics. Third, the top three and top five models with the best performance are utilized as base learners to construct the stacked ensemble deep learning models, and the final classification results are obtained by using majority voting technique. The results show that the stacked integrated network constructed with the top five best-performing basic learners on large-size images yields a peak classification accuracy of 93.98%. In conclusion, the proposed method effectively classifies erythrocyte aggregation levels and elucidates the correlation between image size and the suitable model for this classification task.