Morphology-Guided Muscle Cell Detection and Counting Based on Transfer Learning, FFD Augmentation, and Density-Aware Loss Optimization

In biomedical research, accurate detection and counting of cells in stained soleus muscle sections are crucial for understanding pathological conditions such as muscle atrophy. However, the scarcity of labeled samples and the high cost of manual annotation pose challenges in applying deep learning methods for muscle cell detection and counting. To address these issues, we first employ transfer learning to adapt knowledge from blood cell analysis to muscle cell detection. Second, we simulate muscle stretching or squeezing using Free-form deformation (FFD) and propose a more effective data augmentation strategy. Finally, by leveraging the fact that muscle cells cluster within the same muscle fascicle, we design a density-aware loss function. Using limited microscopic images of rat soleus muscle cells, we demonstrate that our method significantly improves muscle cell detection accuracy.