Interpretable Transfer Learning with EfficientNetB0 for Automated Detection of Rare Anemia Morphologies in Peripheral Blood Smears

Abstract: 

Accurate diagnosis of rare anemias is challenging primarily due to limited annotated public datasets needed for developing interpretable automated diagnostic tools. We propose a two-stage transfer learning convolutional neural network (CNN) model, EfficientNetB0, for identification of disease-associated patterns in blood smear images.

To create and verify the EfficientNetB0 model, we used a two-part proof of concept method. First, model stability was verified using a binary subset of the TensorFlow Flowers dataset (n=1539). It gained a validation accuracy of 0.997 and showed stable learning without overfitting. The validated method was applied to a Sickle Cell Peripheral Blood Smear dataset (n=569), to ensure the model would transfer learning to blood smears. The model was highly successful when trained and validated on the Sickle Cell validation set, with an accuracy of 0.992, precision of 1.000, sensitivity of 0.989, and specificity of 1.000, having no false positives and only one false negative.

Gradient-weighted Class Activation Mapping (Grad-CAM) was used to demonstrate the interpretability of the model, showing that attention was focused on characteristic crescent shaped erythrocytes common in sickle cell blood smears (Acharya & Prakasha, 2019; Elendu et al., 2023), rather than background noise. A MobileNetV2 baseline was trained under identical conditions and achieved significantly lower accuracy and specificity, confirming EfficientNetB0 provides substantially better performance for data scarce classification. These findings show that the EfficientNetB0 transfer learning network is highly accurate, interpretable, and generalizable even with limited data, providing a promising method for automated rare anemia diagnoses.