A calibrated multi-scale transformer-based dual deep convolutional neural network framework for robust parasitic egg recognition across diverse microscopy conditions

The accurate identification of parasitic eggs in microscopic images remains a persistent challenge in both clinical and veterinary diagnostics. Existing manual examination techniques, while effective, are time-consuming, error prone, and difficult to standardize across diverse laboratory settings. Existing automated approaches often struggle to generalize across diverse microscopy conditions and frequently produce poorly calibrated confidence estimates, limiting their practical reliability. To address these limitations, we present calibrated Multi-Scale Transformer Based Dual-Deep Convolutional Neural Network, referred to as MST-Dual-DCNN, for robust parasitic egg recognition across varying microscopy conditions and datasets. The proposed framework combines multi-scale convolutional feature extraction, transformer-based contextual fusion, and dual convolutional neural network backbones (VGG16 and ResNet50) to obtain fine-grained texture and global morphological features. Besides classification, the framework also includes confidence calibration using temperature scaling and produces compact embedding representations to support auxiliary similarity-aware interpretation during inference. A sheep parasite egg dataset was locally developed and contained microscopy samples from three parasite species grouped into two morphologically distinct classification categories, for experiments, along with an independent human hookworm parasite egg dataset for cross-dataset evaluation. The proposed MST-Dual-DCNN framework obtained the best classification accuracy of 98.33%, and an Expected Calibration Error (ECE) of 0.60% post temperature scaling. Extended ablation analysis further showed that multi-scale extraction, transformer fusion and dual-backbone learning improve predictive performance and calibration reliability. The experimental results show that the proposed framework provides accurate and reliable and computationally efficient performance in different microscopy scenarios and also provides improved confidence calibration and classification accuracy compared with the reproduced baseline models. These findings highlight the potential of calibrated hybrid deep learning architectures to assist automated parasitic egg analysis in microscopy-based diagnostic environments.