Optimization of Dropout Rates for Enhanced Generalization in Deep Learning Based Medical Image Segmentation

In recent times, the integration of deep learning into research has surged, leading to the development of more efficient architectures. However, training deep learning models remains a challenge. Despite using similar deep learning architectures, there is no guarantee of equal algorithmic efficiency. One major issue causing reduced efficiency in deep learning algorithms is the lack of diverse training data, often resulting in overfitting. The current literature lacks a precise specification for dropout rates when employed alongside deep learning architectures. This study aimed to determine the optimal dropout rate by experimenting with algorithms designed for segmenting cell images from backgrounds using the U-net deep learning architecture. The findings revealed that there is no definitive numerical value for the ideal dropout rate. Instead, analysis based on the training dataset's characteristics, specifically the diversity of images and the class ratio, provided insights into determining the suitable dropout rate range. Therefore, determining an appropriate dropout rate has become an analytical process reliant on understanding the complexity of the dataset used for training. It involves examining the intricacies of images employed for deep learning and the proportion of classes defined within the training algorithm. This approach allows for a more nuanced and tailored determination of dropout rates, ensuring the efficient functioning of deep learning algorithms in the segmentation of cell images from complex backgrounds.