Intelligence-Based Hematological Diagnostics: Comparative Evaluation of Deep Learning Models for White Blood Cell Segmentation

Proper segmentation of white blood cells (WBCs) plays a crucial role in automated hematological image analysis and computer-assisted diagnostic systems. In this work, the authors compare 5 deep learning architectures: U-Net, U-Net++, Attention U-Net, ResUNet and Mask R-CNN on three datasets BCCD, Raabin-WBC and TNBC-WBC. A ten-fold cross-validation was performed for each dataset separately, with a shared data-augmentation and preprocessing scheme. The Dice coefficient, Intersection over Union, precision and recall were used to evaluate the segmentation performance. The computational properties such as model size, floating-point operations, training time and inference time were also investigated. To understand the regions in the image that affect model predictions, qualitative interpretability analysis was conducted using Grad-CAM and the attention-gate visualizations. The highest cross-dataset macro-average performance was obtained by U-Net++ with the Dice coefficient of 94.27% and the IoU of 89.50%. U-Net and ResUNet also showed competitive results, especially on the images that are heterogeneously stained and have low contrast and complex cellular morphology. The architecture-specific segregation trade-offs between performance and complexity of the model and inference speed were identified using computational profiling. Overall, the results indicate that the U-Net++ is the best of the models analyzed, and Attention U-Net and ResUNet are competitive for the difficult WBC images. The study offers a uniform benchmarking platform for assessing performance, computational complexity, and interpretability of deep learning models for WBC segmentation.

Hali tarjima qilinmagan