We are pleased to announce that our latest research has been published in Life (MDPI, Open Access).
The study, titled “Adaptive Logit Fusion for Mitigating Class Imbalance in Multi-Category Sperm Morphology Assessment,” presents a deep learning approach that classifies sperm cells into 18 morphological classes (one normal and 17 abnormal types). Two convolutional networks, EfficientNetV2-S and ResNet50V2, are fine-tuned with a class-weighted loss and extensive data augmentation, and an ensemble is then built by linearly fusing the logits of both architectures, with the fusion weight optimized to maximize recall, precision and F1-score.
On the heavily imbalanced Hi-LabSpermMorpho data, the proposed ensemble reaches an overall accuracy of 70.94%, consistently outperforming the individual models. Cells with pronounced structural abnormalities such as PinHead and DoubleTail are classified with high accuracy, whereas visually subtle defects remain more challenging, pointing to where future work on automated sperm morphology assessment should focus.
Authors:
Emin Can Özge (Siemens A.Ş. / Yıldız Technical University)
Hamza Osman İlhan (Yıldız Technical University)
Görkem Serbes (Yıldız Technical University)
Hakkı Uzun (Recep Tayyip Erdoğan University)
Ali Can Karaca (Yıldız Technical University)
Merve Hüner Yiğit (Recep Tayyip Erdoğan University)
