Unsupervised image classification with patch shared hierarchical SOM on MNIST and for clinical blastocyst quality assessment

This study introduces PS-HSOM, a patch-shared hierarchical Self-Organizing Map, and evaluates its unsupervised representation learning performance through post hoc cluster-to-label evaluation across diverse application domains. When applied to the MNIST dataset, the model attains 90.5% test accuracy after post hoc cluster-to-label mapping using only 4000 training samples and three training epochs per layer. This result exceeds the performance of E-DSOM by 3.4%, while requiring approximately 1.33 × more training data, but approximately two orders of magnitude fewer parameters, and offering full experimental reproducibility through fixed random seed initialization. The framework is further validated on a clinical embryo assessment task using the Linz IVF dataset, comprising 1615 training images and 214 test images. In this setting, PS-HSOM achieves 97.2% accuracy in blastocyst expansion stage (EXP) classification without employing morphological labels during feature learning. These findings suggest that PS-HSOM can achieve clinically promising performance for unsupervised embryo grading under limited-data conditions.