The Effect of Class-Weighted Penalization in Deep Neural Networks for Multi-Class Cell Segmentation

Deep learning networks give successful results in many areas, but their complexity leads to problems such as overfitting. Many approaches have been proposed to solve this problem, and class-based penalization has been one of the methods that have yielded successful results. With class-based penalization, it has become possible to increase the prediction performance and improve the model's generalization capability, especially in cases with class imbalance. This study investigates the effect of class-based penalization on the multiclass cell segmentation problem. Two deep neural network models (Resnet18, EfficientNet) are tested with 6 different configurations created for class-based penalization, and the results are compared. The experimental studies show the relationship between class-based loss penalties and multiclass segmentation/classification performance. The results show that class-based penalization improves the total performance of EfficientNet and Resnet18 networks by 11.82(C2+C3) and 12.79(C2+C3) points, respectively. It is shown that the proposed method can improve the prediction performance without increasing the model complexity.