We are pleased to announce that our latest research has been published in Diagnostics (MDPI, Open Access).
The study, titled “A Multi-Teacher Knowledge Distillation Framework for Enhancing the Robustness of Automated Sperm Morphology Assessment,” proposes knowledge distillation as a strong regularizer for sperm morphology classification on small, heavily imbalanced clinical datasets. Soft distillation transfers knowledge from high-capacity teachers (SwinV2-large, EfficientNetV2-m, ConvNeXtV2-large) to a smaller SwinV2-base student, comparing single-teacher and multi-teacher (averaged-response) strategies.
Using a cross-dataset protocol on the 18-class Hi-LabSpermMorpho dataset — teachers fine-tuned on two staining sets (BesLab, Histoplus, GBL) and the student trained on the third — the multi-teacher student with augmentation and soft distillation reached 70.94% (BesLab), 73.61% (Histoplus) and 71.63% (GBL) accuracy, consistently above the baselines, offering a more generalizable solution for clinical sperm morphology assessment.
Authors:
Osman Emre Tutay (Yıldız Technical University)
Hamza Osman İlhan (Yıldız Technical University)
Hakkı Uzun (Recep Tayyip Erdoğan University)
Merve Hüner Yiğit (Recep Tayyip Erdoğan University)
Görkem Serbes (Yıldız Technical University)
