A deep learning-based U-Net 3+ technique for segmentation blood cell

Abstract

This study introduces a novel and enhanced UNet3Plus model tailored for the precise segmentation of blood cells in medical images. The architecture incorporates structural modifications, including strengthened connections between convolutional layers, increased filter numbers, and integration of Bayesian optimization for hyperparameter tuning. The model's generalization capability is optimized through the dynamic adjustment of dropout rates and learning rates. Bayesian optimization facilitates the exploration of optimal hyperparameter combinations, allowing the model to adapt effectively to diverse datasets. Advanced training strategies, such as adaptive learning rate adjustment and early stopping, are employed to mitigate overfitting and enhance training efficiency. The proposed model exhibits exceptional performance across multiple folds, achieving low training and validation losses, high accuracy metrics, and robust segmentation indices. Evaluation metrics, including Mean IoU (Jaccard Index), Dice score, Pixel Accuracy, and Precision, affirm the model's proficiency in accurately delineating blood cell boundaries. The study contributes to the field of deep learning-based medical image segmentation by showcasing the effectiveness of customized architectures and optimization techniques. The proposed UNet3Plus model stands as a promising solution for accurate and reliable blood cell segmentation, demonstrating adaptability and robust performance across various datasets. This work sets the stage for future research in the domain of medical image segmentation, emphasizing the potential for continued advancements in precise and efficient segmentation methodologies.