algorithm, Image processing.

Automatic analysis of microscopic blood images has become one of the important research topics in the field of medical image processing and Computer-Aided Diagnosis (CAD) systems over the past two decades. Differential White Blood Cell Count (DWBCC) is one of the most common hematological tests that provides valuable information regarding the immune system status, various infections, inflammatory diseases, autoimmune disorders, and hematological malignancies such as leukemia. However, performing this test manually is not only time-consuming but also highly dependent on the expertise of specialists and may be affected by fatigue, inter-observer variability, and human errors. Therefore, the development of intelligent systems for the automated analysis of blood smear images has emerged as an active research area in medical image processing [1–3]. In most automated hematological disease diagnosis systems, White Blood Cell (WBC) segmentation is considered the first and one of the most critical processing stages, as the accuracy of subsequent steps, including feature extraction, cell classification, and disease diagnosis, directly depends on the quality of segmentation. Any error in identifying cell boundaries or incomplete extraction of the nucleus and cytoplasm regions may lead to inaccurate extracted features and consequently reduce the performance of classification models. Therefore, accurate WBC segmentation has been recognized by many researchers as one of the fundamental challenges in automated blood image analysis systems [2], [4]. Although the detection of WBCs in stained microscopic images may appear straightforward, this process is practically challenging due to several factors. Variations in staining protocols (such as Wright, Giemsa, and Wright–Giemsa), differences in imaging conditions, illumination variations, noise, color deviations caused by laboratory equipment, overlapping cells, the presence of platelets and artifacts, variations in WBC morphology and size, as well as differences among available datasets, cause many segmentation algorithms to perform effectively only under specific conditions and exhibit limited generalization capability when applied to new images [3], [5].

To overcome these challenges, various approaches have been proposed for WBC segmentation. The early generation of these methods was mainly based on intensity thresholding and color information. In these approaches, the color difference between the WBC nucleus and other image components is exploited to separate WBCs from the background and Red Blood Cells (RBCs). Although these methods are computationally simple and fast, their performance is highly dependent on threshold selection and they demonstrate limited robustness against illumination variations, staining differences, and image quality changes [6].

To improve segmentation accuracy, researchers have investigated the use of different color spaces, including HSV, Lab, YCbCr, and CMYK. The main idea behind these approaches is that certain color channels can enhance the contrast between WBC nuclei and other image components, thereby facilitating the segmentation process. However, selecting an appropriate color space remains dependent on dataset characteristics, staining protocols, and imaging conditions, and no single-color space has demonstrated optimal performance across all datasets [5], [7].

In recent years, with the advancement of machine learning techniques, particularly deep learning methods, numerous studies have been conducted for WBC segmentation and classification. Although these approaches have achieved promising performance on several benchmark datasets, recent systematic reviews indicate that challenges such as dependency on image quality, the requirement for large annotated datasets, limited generalization capability, and high computational costs remain unresolved [8], [9].

Alongside deep learning-based approaches, classical image processing techniques continue to be widely employed in blood image analysis systems due to their simplicity, interpretability, computational efficiency, and independence from training procedures. Among these techniques, thresholding methods, morphological operations, distance transform, region growing, and Marker-Controlled Watershed algorithms have been extensively utilized [2], [10]. These methods can accurately extract WBC nuclei and cell regions under controlled imaging conditions; however, their performance is strongly influenced by marker selection, staining uniformity, and the degree of cell overlap.

To reduce the over-segmentation problem, several modified versions of the Watershed algorithm have been proposed. For instance, Abrol et al. developed an improved Marker-Controlled Watershed approach to enhance leukocyte separation accuracy; however, the selection of initial markers remained dependent on image characteristics [10]. Furthermore, several studies have demonstrated that the performance of Watershed-based algorithms significantly decreases in images with severe illumination variations, non-uniform staining, or clustered cells [2], [8].

In recent years, Convolutional Neural Networks (CNNs), U-Net architectures, Attention U-Net, Mask R-CNN, and other deep learning models have been extensively applied for WBC segmentation. These approaches achieve high accuracy on public datasets by automatically learning complex image features [9], [11]. Nevertheless, most of these models require large amounts of annotated data, considerable training time, powerful computational hardware, and careful hyperparameter optimization. Moreover, recent investigations have shown that the performance of many deep learning models decreases when evaluated on images acquired from different laboratories, imaging devices, or staining protocols, indicating that generalization remains a major challenge in this field [8], [9].

A review of previous studies reveals that most existing methods have focused on improving a specific stage of the segmentation process; for example, some studies emphasize accurate nucleus extraction, others focus on separating overlapping cells, and some aim to enhance classification performance [2], [8], [11]. In contrast, only a limited number of studies have attempted to develop an integrated and imaging-condition-independent framework that simultaneously exploits stable characteristics of blood smear images for background extraction, nucleus detection, RBC removal, adjacent cell separation, and false region elimination. This issue has been highlighted as an important research gap in recent reviews [8], [9]. Although numerous WBC segmentation methods have been introduced in recent years, achieving a robust and generalizable approach capable of providing stable and reliable performance under diverse imaging conditions and across different datasets remains an open challenge [3], [8], [9]. Many classical methods, including thresholding, morphological operations, clustering, and Watershed-based techniques, have been specifically optimized for particular datasets or imaging conditions and require parameter adjustment when staining type, illumination intensity, image quality, or cell overlap characteristics change [2], [4], [10]. On the other hand, although deep learning-based methods have reported high accuracy on training datasets, their dependence on large annotated datasets, high computational requirements, and reduced generalization ability when applied to images from different laboratories and imaging systems remain significant limitations [8], [9], [11].

Previous studies indicate that despite considerable progress, developing an accurate and stable segmentation method that maintains consistent performance under different imaging conditions and across diverse datasets remains a significant challenge. Many classical approaches focus on improving a specific segmentation stage, such as nucleus extraction or separation of adjacent cells, and require parameter recalibration when image quality or staining conditions vary [2], [5]. Conversely, although deep learning approaches have demonstrated high accuracy on certain datasets, their dependence on extensive training data, computational complexity, and limited transferability to images acquired from different laboratories restrict their practical applicability [6–8]. Furthermore, in many existing studies, different segmentation stages are performed independently, and the informative relationship between background extraction, nucleus detection, RBC removal, and adjacent cell separation is not fully exploited. Consequently, errors generated in early stages may propagate to subsequent steps and negatively affect the final segmentation accuracy.

In this study, a sequential framework for WBC segmentation in peripheral blood smear images is proposed, where the output of each stage is utilized as guiding information for the subsequent stage. In the proposed approach, the image background is initially extracted using the M channel of the CMYK color space and the Zack thresholding algorithm. Subsequently, WBC nuclei are detected, and based on the information obtained from these two stages, RBCs are separated from other image regions. Then, initial WBC regions are extracted, and an improved version of the Marker-Controlled Watershed algorithm based on distance transform is employed to separate adjacent cells. Finally, false regions are removed according to geometric characteristics, generating the final WBC segmentation mask.

The main contributions of this study are summarized as follows:

I. A sequential and integrated framework is introduced for WBC segmentation, in which the extracted information from each stage is utilized as guidance for the following stage.

II. The combination of background, nucleus, and RBC information is employed to improve the WBC segmentation process.

III. An improved Marker-Controlled Watershed algorithm is proposed to enhance the separation of clustered and adjacent WBCs.

IV. The proposed method is evaluated on multiple standard datasets to investigate its accuracy, robustness, and generalization capability.

Microscopic blood smear images are generally acquired as color images and contain four meaningful regions, including the WBC nucleus, WBC cytoplasm, RBCs, and background. Although the color characteristics of these regions may vary depending on imaging conditions and staining protocols, three common assumptions exist for microscopic blood smear images that remain invariant under different acquisition conditions and staining variations. These assumptions are described as follows:

Assumption 1: The darkest purple region in microscopic blood smear images corresponds to the nuclei of WBCs.

Assumption 2: The brightest region in these images represents the background area.

Assumption 3: No nucleus exists within the regions corresponding to RBCs in microscopic blood smear images.

Due to the invariance of these three assumptions against changes in imaging conditions and staining procedures, they are employed in this study to design a robust framework for WBC segmentation. In this context, robustness refers to the capability of the proposed segmentation method to accurately segment WBC regions across various images acquired from different laboratories with diverse staining protocols and imaging systems, regardless of the specific image characteristics. An algorithm can be considered robust when both its overall structure, the interaction between its different components, and the employed techniques in each processing stage maintain stable performance under varying conditions. In the proposed WBC segmentation algorithm, the aforementioned three assumptions are utilized to construct a robust framework. Furthermore, at each stage of the algorithm, reliable and robust methods available in the corresponding field are employed to enhance the stability and generalization capability of the segmentation process. Overall, the proposed segmentation algorithm consists of six main stages. Fig. 1 illustrates the block diagram of the proposed framework, and each stage is described in detail in the following sections.

In this paper, images from three different datasets with varying characteristics were employed to evaluate the different stages of the proposed method. Details of each dataset are described below.

Dataset 1: This dataset consists of 400 images acquired from 100 peripheral blood smears. The blood smears were stained using the Giemsa-Wright method. The images were captured by a digital camera (SONY model No. SSC-DC50AP) mounted on an optical microscope (Microscope-Axioskope 40) equipped with a colorless lens at 100× magnification. All images are color images of size 720 × 576 pixels with 24 bits per pixel, stored in BMP format. The images were collected at the Hematology-Oncology Research Center of Imam Khomeini Hospital (Tehran, Iran) and classified into different types of WBCs by a hematologist. In addition, the nuclear and cytoplasmic regions of the WBCs were manually segmented by an expert. The images in this dataset exhibit significant variation in brightness intensity. This dataset is suitable for nucleus and cytoplasm segmentation of WBCs as well as for the classification of different types of leukocytes [15].

Dataset 2: The microscopic images in this dataset were acquired using a KODAK EasyShare V803 digital camera mounted on a Leica microscope at 100× magnification with an oil-immersion lens. The required blood smears were stained with Giemsa. The images are color images of size 3264 × 2448 pixels with 24 bits per pixel, stored in JPG format. These images were collected by the specialized Hematology Center at Dr. Juan Bruno Zayas Alfonso Hospital in Santiago, Cuba. Healthy and pathological RBCs were classified by a specialist in this dataset. This database can be used for blood smear image segmentation and RBC classification methods.

Dataset 3: The images in this dataset were captured using a CANON Power Shot G5 digital camera attached to a laboratory microscope with variable magnification (ranging from 300× to 500×). The images are color images with 24 bits per pixel, stored in JPG format. This dataset was prepared at the M. Tettamanti Research Center in Monza, Italy, for the detection of Acute Lymphoblastic Leukemia (ALL) using blood sample image processing techniques. It consists of two subsets: ALL-IDB1 and ALL-IDB2.

The first subset (ALL-IDB1) contains 108 images of size 2592 × 1949 pixels, encompassing a total of 3900 blood elements. Overall, there are 510 lymphoblasts across all images, whose center coordinates were determined by an expert. In each image, the presence or absence of at least one lymphoblast is also indicated. The number of images containing lymphoblasts is equal to the number of images without lymphoblasts. This subset can be used to evaluate the algorithm’s performance in blood smear segmentation and separation of overlapping cells. Moreover, due to the availability of labeled lymphoblasts, it is suitable for assessing the classification power of the system in ALL detection at both image and cell levels.

Fig. 6 presents the practical implementation of the first six stages of the proposed WBC segmentation algorithm on a sample microscopic blood smear image.

generated label map, e. red blood cell regions, f. candidate white blood cell regions, g. segmented white blood cell regions, and h. final white blood cell regions after removing WBC-like objects.

Figs. 7-9 illustrate the segmentation results obtained by the proposed algorithm on images from different databases.

As illustrated in Figs. 7–9, the proposed framework demonstrates stable segmentation performance across all three datasets, despite considerable variations in image acquisition conditions, staining protocols, image quality, and imaging devices. In all test images, the initial extraction of the background and WBC nuclei provides reliable information for the subsequent processing stages. The sequential utilization of this information effectively reduces error propagation throughout the segmentation pipeline.

Furthermore, the experimental results indicate that exploiting RBC information before the final WBC extraction substantially eliminates irrelevant regions and improves the accuracy of candidate WBC detection. In addition, the modified Marker-Controlled Watershed algorithm based on the distance transform successfully separates clustered WBCs, even in challenging cases involving extensive cell overlap, while reducing the over-segmentation commonly observed in conventional Watershed-based approaches.

Overall, the results obtained from the three independent datasets demonstrate that the proposed framework maintains stable performance under different imaging and staining conditions without requiring dataset-specific parameter adjustment. These findings indicate that the proposed method possesses satisfactory robustness and generalization capability, making it suitable for automated peripheral blood smear image analysis.

In this study, a robust multi-stage framework for WBC segmentation in microscopic peripheral blood smear images was proposed. Unlike many existing approaches whose performance depends heavily on imaging conditions, staining protocols, or specific datasets, the proposed framework was developed based on three invariant characteristics of peripheral blood images. Furthermore, the output of each processing stage was utilized as guiding information for the subsequent stage, resulting in a coherent segmentation pipeline that minimizes error propagation and improves robustness under varying imaging conditions.

The proposed framework integrates background extraction using the M channel of the CMYK color space and the Zack thresholding algorithm, WBC nucleus detection, RBC region extraction, preliminary WBC segmentation, separation of clustered cells using a modified Marker-Controlled Watershed algorithm based on the distance transform, and the removal of WBC-like artifacts using geometric features within a unified segmentation strategy.

Experimental evaluation on three publicly available blood smear datasets acquired under different imaging and staining conditions demonstrated that the proposed method provides accurate and stable WBC segmentation without requiring dataset-specific parameter adjustment. The results further indicate that exploiting complementary information from background extraction, nucleus detection, and RBC identification, together with the improved marker generation strategy for the Marker-Controlled Watershed algorithm, significantly enhances the separation of adjacent WBCs while effectively reducing over-segmentation. Overall, the proposed method offers an interpretable, computationally efficient, and robust alternative to existing segmentation approaches and is well suited for automated blood image analysis systems. Future work will focus on integrating the proposed framework with deep learning and transfer learning techniques to further improve segmentation performance on highly challenging blood smear images while preserving the robustness and interpretability of the proposed methodology.

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

All data are included in the text.

The authors declare no conflict of interest.