1 Introduction

As diseases in humans are gradually rising, an automated disease detection system is critically needed, especially when a mass screening process is essential [1,2,3,4]. Disease diagnosis with biomedical imaging is often required, in which the disease can be detected with the help of a chosen imaging modality [5, 6]. In the most chronic and infectious disease screening procedures, blood screening is a mandatory procedure, where a blood sample is collected from the patient. Detection of leukocyte (white blood cell) is particularly an essential practice during the screening of infectious diseases. This is normally performed using blood smear images (thin/thick) collected by digital microscopes [7]. Normally, hematological images collected using prescribed clinical protocol show vital information about the health condition of the patient. Further, assessment of the leukocyte count per micro-liter is also an approved practice to verify the health condition and to check the immunity level.

This work aims to develop an automated segmentation system to extract various classes of the leukocyte images available in the clinical-grade hematological test images. In the literature, a number of semi-automated/automated image processing actions have been proposed, which examined leukocyte from thin blood smear images. Our work aims to use the Convolutional Neural Network (CNN) supported segmented scheme to extract the leukocyte section with better segmentation accuracy. In the literature, a number of pre-trained and customary CNN segmentation schemes are available, which examine biomedical images recorded using varied imaging modalities [8,9,10,11,12]. The development of a customary CNN scheme for a chosen image is computationally complex, and hence, pre-trained CNN designs are extensively adopted by most researchers due to its availability, performance, and adaptability toward varied imaging modalities [13,14,15,16,17].

In this research, we adopted well-known CNN schemes, such as SegNet [9, 10], U-Net [11,12,13], and VGG-UNet [14,15,16,17], to extract leukocyte fragments from images with enhanced accuracy. After extracting the leukocyte fragment, a relative assessment between the extracted section and the available Ground-Truth (GT) is done to confirm the performance of the CNN scheme. The test images used in this proposed research are collected from Leukocyte Images for Segmentation and Classification (LISC) dataset [18].

This dataset consists of the five classes of leukocyte images with a dimension of \(720x576x3\) pixels. LISC database comprised of 376 images in which 250 images are available with GT, and the remaining 126 images are without the GT. During the experimental evaluation, each image is resized into \(256x256x3\) pixels and then the image augmentation is implemented during the training of the CNN. After the training, the performance of the CNN segmentation scheme is tested using the 250 images, which are available with the GT. The experimental investigation is separately implemented with SegNet, U-Net, and VGG-UNet. The extracted leukocyte fragment is then compared with the GT. We also calculate the Image Performance Measures (IPM), and based on these values, the performance of the CNN segmentation schemes is validated. The experimental outcome confirms that the CNN scheme is a promising automated segmentation technique. After pre-tuning, the VGG-UNet scheme offered a better outcome compared to SegNet and U-Net. The performance of the proposed scheme is then confirmed and validated using similar medical images existing in the literature, such as Blood Cell Count and Detection (BCCD) database and ALL-IDB2. For every image cases, the proposed scheme helps to get better segmentation accuracy and this confirms that the proposed scheme works well for hematological images. The main contribution of this research work is: (1) Employing the pre-trained CNN segmentation technique to extract the abnormal section from the test image and (2) Performance evaluation of the proposed scheme on well-known benchmark images.

The remaining sections of this work are structured as follows: Sect. 2 describes the related work, Sect. 3 explains the adopted methodology, Sects. 4 and 5 present the results and conclusion of this research, respectively.

2 Related work

Blood screening is a commonly adopted disease prescreening procedure, and the assessment of leukocyte type and its count is also an essential practice in clinical level assessment. Due to its significance, a considerable number of leukocyte examination procedures are proposed and implemented in the literature. Table 1 presents a summary of various image processing procedures implemented to examine the hematological images.

Table 1 Summary of recent leukocyte segmentation techniques for hematological images
Full size table

Table 1 presents the few recently implemented leukocyte segmentation technique using traditional and DNN-based techniques. A short review of the segmentation technique employed to extract the White Blood Cell (WBC) can be found in the work of Sapna and Renuka [31]. All this work confirms that the extraction and evaluation of the WBC is a significant task during the blood level disease detection and to reduce the diagnostic burden, it is necessary to employ an automated WBC evaluation system. The recent works in the literature confirms that the CNN approaches help to achieve a superior result during the data assessment [32,33,34,35]. Hence, this research aims in implementing the CNN supported scheme to assess the considered database. An automated hematological image examination scheme should have the capability to detect/classify the leukocyte (WBC) with better accuracy. To achieve an automated detection, this research work employed the pre-trained CNN scheme available in the literature, and the performance of the considered CNN network is then confirmed with an experimental study using the benchmark LISC database.

3 Methodology

This section describes the proposed scheme, the image database used as a benchmark, the adopted CNN schemes, as well as the performance measures.

3.1 Proposed scheme

The structure of the proposed scheme and its stages are depicted in Fig. 1. Initially, the essential test images are gathered from the LISC database. All the RGB images of the LISC database are resized to \(256x256x3\) pixels to reduce the computation complexity. The existing pre-trained CNN schemes, such as SegNet, U-Net, and VGG-UNet, are then used to extract the important fragment from the test images. The chosen CNN scheme is trained using the existing LISC dataset images and during this task, image augmentation, such as flip and rotate is used to increase the learning capability of the CNN. After the training process is completed, the original LISC images along with the GT (250 images) are then used to test the performance of the CNN. After image segmentation, the binary version of the segmented image is considered as the outcome and finally, this image is compared against its related GT. The results from the CNN are also compared with results from the existing literatures.

Fig. 1
figure 1

Structure of the proposed CNN segmentation scheme

Full size image

3.2 Image database

The development of an appropriate disease detection system is essential in the medical domain. Validating the system with the clinical-grade benchmark images is also critical. In this research, we use the LISC dataset, which is one of the clinical-grade leukocyte image datasets, developed in the year 2010 by Rezatofighi et al. [19]. This dataset consists of five categories of images, namely basophil, eosinophil, lymphocyte, monocyte, and neutrophil, as well as mixed cases. All these images are associated with the GT, and this dataset also has 126 images without the GT. Other related information, such as patient-related details, can be accessed from [18,19,20]. Figure 2 illustrates a sample test imagery of LISC with different classes.

Fig. 2
figure 2

Sample test images available in the LISC database

Full size image

Along with the LISC images, this research work also considers Blood Cell Count and Detection (BCCD) database [36] and ALL-IDB2 [37,38,39] to test the performance of the proposed scheme and the sample images from this database are depicted in Fig. 3. In this work, 250 images from each database is considered to validate the performance of the employed segmentation system.

Fig. 3
figure 3

Sample images of Leukocyte data considered for validation

Full size image

3.3 CNN scheme

During the medical image diagnosis, the commonly performed image processing procedures are segmentation and classification. During the segmentation task, the essential image section (Region-Of-Interest) is extracted using a chosen technique and is then evaluated using a chosen computer algorithm to detect the disease. The development of accurate image segmentation is always essential to get better disease detection accuracy.

In the literature, a significant amount of traditional [30, 40, 41] and modern (CNN) [9, 11, 42, 43] medical segmentation techniques have been used. Implementation of the traditional segmentation techniques is time consuming, and most of the existing traditional techniques are semi-automated methods and frequently need operator assistance. Due to this reason, modern techniques are widely preferred to examine medical images of varied modalities. Recently, pre-trained CNN schemes have been extensively adopted in the image processing domain, in which the pre-trained CNN schemes work well on a class of images with varied dimensions. Further, the trained CNN on a particular image case will produce better results compared to the traditional approaches. In this work, the CNN schemes, such as SegNet, U-Net, and VGG-UNet are considered to examine the LISC images. These CNN schemes are initially trained with the images of the LISC. During this process, the original as well as the augmented images are considered. After the training, the performance of the CNN is tested and validated using the leukocyte images available with the GT.

3.3.1 SegNet

SegNet is a well-known CNN scheme proposed in 2015. This scheme is widely used to implement the pixel-wise analysis of RGB/gray-scale images [9,10,11].

The SegNet is constructed by implementing a series of structured Convolutional Encoder-Decoder (CED) framework, and every framework transfers the learned information to the next successive section. The structure of the traditional SegNet is depicted in Fig. 4. In the implementation, we used the following parameters: image augmentation is fixed as linear, learning rate is assigned as 0.005, decoder-encoder batch size is fixed as 4, normal weight initialization, linear dropout rate and Stochastic Gradient Descent (SGD) adaptive learning rate is considered. The last layer of this scheme is equipped with a Sigmoid activation that provides a classified binary image (which groups the pixel into two groups, such as the Leukocyte section and background). The final outcome of the SegNet is converted into a binary image in order to compare it with the binary GT.

Fig. 4
figure 4

Structure of the SegNet scheme

Full size image

3.3.2 U-Net

U-Net was proposed in 2015 as a sliding window convolutional network, dedicatedly developed to examine test images of the ISIC challenge database [12]. In recent years, due to its performance and significance, a considerable number of modified versions of U-Net schemes are available for other image databases [13,14,15,16].

The U-Net scheme used in this research is adopted from the work of El Adoui et al. [11], and the architecture is presented in Fig. 5. Here, the test image and the segmented image have the dimension of \(256x256x3\) pixels. The initial tuning of the U-Net is the same as those for the SegNet, and are similar for the VGG-UNet scheme. The working methodology is also similar to the conventional encoder-decoder scheme. Finally, the classifier unit helps to get the outcome with two class image pixels grouped as the binary image. Other related information on the conventional U-Net can be found in [13,14,15,16,17].

Fig. 5
figure 5

Structure of the U-Net scheme

Full size image

3.3.3 VGG-UNet

A substantial number of CNN segmentation schemes are available in the literature, and the VGG-UNet is one of the enhanced forms of the U-Net scheme. The working parameters and the pre-tuning of the VGG-UNet are similar to the U-Net, and in this approach, the learned features of the VGG16 scheme is considered to enhance the segmentation accuracy. During the implementation, all the examination images are resized into \(224x224x3\) pixels. The binary image produced by this scheme is resized into \(256x256x1\) to have a fair comparison with the other CNN segmentation methods. In this work, the Convolutional part of the VGG16 will act as the encoder part and the Up-Convolutional part of the U-Net act as the decoder part. Finally, the Sigmoid activation helps to get the segmented result. Other information related to the VGG-UNet can be found in the following work [13,14,15,16,17].

3.4 Performance measures

CNN segmentation performance needs to be authenticated by calculating the image performance values. After extracting the leukocyte segment from the chosen hematological image, a relative assessment with the existing GT is then performed and the essential values of the performance measures, such as Jaccard-Index (JI), Dice-Coefficient (DC), Accuracy (AC), Precision (PR), Sensitivity (SE), Specificity (SP), and Negative-Predicted-Value (NPV), are calculated. Based on these values, the performance of the SegNet, U-Net, and VGG-UNet is validated. This comparison uses the binary images, in which the leukocyte region is considered as Positive (P) pixel (binary1) and the background section is accounted as Negative (N) pixel (binary0). This comparison helps to compute the measures depicted in Eqs. (1) to (9) [45,46,47,48].

$$ FP_{{{\text{rate}}}} = FP/N = {{FP} \mathord{\left/ {\vphantom {{FP} {(TN + FP)}}} \right. \kern-\nulldelimiterspace} {(TN + FP)}} $$
(1)
$$ FN_{{{\text{rate}}}} = FN/P = {{FN} \mathord{\left/ {\vphantom {{FN} {(TP + FN}}} \right. \kern-\nulldelimiterspace} {(TP + FN}}) $$
(2)
$$ JI = {{F1 - Score = TP} \mathord{\left/ {\vphantom {{F1 - Score = TP} {(TP + FP + FN}}} \right. \kern-\nulldelimiterspace} {(TP + FP + FN}}) $$
(3)
$$ DC = {{2TP} \mathord{\left/ {\vphantom {{2TP} {(2TP + FP + FN}}} \right. \kern-\nulldelimiterspace} {(2TP + FP + FN}}) $$
(4)
$$ AC = (TP + TN)/(TP + TN + FP + FN) $$
(5)
$$ PR = {{TP} \mathord{\left/ {\vphantom {{TP} {(TP + FP)}}} \right. \kern-\nulldelimiterspace} {(TP + FP)}} $$
(6)
$$ SE = TP_{{{\text{rate}}}} = TP/P = {{TP} \mathord{\left/ {\vphantom {{TP} {(TP + FN}}} \right. \kern-\nulldelimiterspace} {(TP + FN}} $$
(7)
$$ SP = TN_{{{\text{rate}}}} = TN/N = {{TN} \mathord{\left/ {\vphantom {{TN} {{\text{(TN}} + {\text{FP)}}}}} \right. \kern-\nulldelimiterspace} {{\text{(TN}} + {\text{FP)}}}} $$
(8)
$$ NPV = {{TN} \mathord{\left/ {\vphantom {{TN} {(TN + FN)}}} \right. \kern-\nulldelimiterspace} {(TN + FN)}} $$
(9)

where TN, TP, FN, and FP represent true-negative, true-positive, false-negative, and false-positive, respectively.

4 Results and discussion

After image resizing and CNN pre-tuning with the LISC dataset, the proposed segmentation is initially implemented using the test images with the GT. Figure 6 depicts the sample test image (Basophil class). Figure 6a, b represent the resized examination image and the related GT, respectively. Figure 6c depicts the saliency map generated by CNN during the learning process, and Fig. 6d shows the extracted binary image with the SegNet scheme. The saliency map clearly shows that CNN precisely identified the section (leukocyte) to be extracted by the final Sigmoid activation function. A similar procedure is then repeated using U-Net and VGG-UNet schemes. Their results are depicted in Fig. 6e, f, respectively. From these images, it is clearly seen that, when the CNN is perfectly trained with the considered image database, it identifies and segments the leukocyte section with better accuracy.

Fig. 6
figure 6

Sample results using Basophil class test image Test image, b GT, c Saliency map, d–f shows extracted Leukocyte image using SegNet, U-Net, and VGG-UNet, respectively

Full size image

After mining the leukocyte, a number of measurements using Eqs. (1) to (9) are computed. The comparison results between Fig. 6b, d–f are shown in Tables 2 and 3. The performance measure obtained by SegNet is shown to be better than those by U-Net and the VGG-UNet. The process is then replicated to all LISC images.

Table 2 Performance values for the extracted leukocyte with GT
Full size table
Table 3 Essential performance values computed using leukocyte and GT comparison
Full size table

The LISC dataset also consists of some complex hematological images which are challenging to many proposed computer-assisted disease detection tools, as these images are associated with more than one leukocyte section. Figure 7 shows the images with two and three leukocyte sections in a single image frame.

Fig. 7
figure 7

Results for complex blood smear images

Full size image

Figure 6 shows the image and the GT, as well as the saliency map by the SegNet scheme. The saliency map clearly confirms the correctness of the pre-training procedure implemented on the CNN segmentation using the LISC dataset. Due to its initial training, the CNN architecture will remember the pixel groups belong to the leukocyte section and efficiently recognizes all the pixels to be extracted. From the saliency map, it is clear that the enhanced pixels belong to the leukocyte section, which will be identified and extracted by the final pixel classification layer. Approximately, similar results are obtained by both U-Net and VGG-UNet.

CNN segmentation technique is applied to the entire LISC dataset, which has the GT. The results for each leukocyte class are presented in Table 4. From this table, it is clear that SegNet, U-Net, and VGG-UNet successfully produce good image performance measures when compared with the GT, and this demonstrates that the abilities of CNN segmentation schemes.

Table 4 Performance measures of individual leukocyte image class
Full size table

To identify the overall performance measure, the average performance measure for each CNN scheme is separately computed. The overall measure is then compared using the Glyph-Plot [36] as shown in Fig. 8. This figure verifies that the overall performance by VGG-UNet is better compared to SegNet and U-Net. Compared to the SegNet, the traditional U-Net showed poor performance. This performance can be improved by varying the initial parameters of the pre-trained U-Net architecture.

Fig. 8
figure 8

Glyph-plot demonstrating the overall performance measure

Full size image

The results obtained from this study confirm that the CNN schemes are automated schemes and work well on the leukocyte images with varied classes. Further, the overall performance measures, such as JI, DC, AC, PR, SE, SP, and NPV show promising results on each segmentation scheme.

The results of the proposed research are also evaluated against the existing semi-automated and hybrid image segmentation procedures available in the literature, and the results are presented in Figs. 9 and 10.

Fig. 9
figure 9

Glyph-Plot of the overall performance measures obtained with Basophil image class

Full size image
Fig. 10
figure 10

Performance evaluation between VGG-UNet with existing semi-automated and hybrid image processing schemes

Full size image

Figure 9 depicts the performance evaluation with Basophil image class and the results obtained by the Chan-Vese segmentation [28] technique and the CNN scheme. The Chan-Vese segmentation is used after the image thresholding process. The overall performance measure is approximately similar to the results of VGG-UNet.

Figure 10 presents a comparison result between the earlier works performed on the LISC database, such as Chan-Vese [28], Level set [29], and Hough transform [30] with proposed VGG-UNet. The earlier works discussed in the literature [28–30] are hybrid image processing methods, in which the preprocessing is performed using a heuristic algorithm-assisted thresholding procedure, and mining is employed with the chosen segmentation technique. The overall accuracy by the VGG-UNet is quite similar to that of the existing methods, and sensitivity and specificity are better compared to the existing traditional procedures.

Figure 11 presents the sample results achieved with VGG-UNet, and this result confirms that the outcome in both BCCD and ALL-IDB2 database is good. This result is compared with the segmentation result of VGG-SegNet, and the outcome is presented in Fig. 12. From this figure, it can be noted that, the VGG-UNet based segmentation helps to get a better average values of JI, DC, and AC compared to UNet, SegNet, and VGG-SegNet. Figure 12a presents the BCCD database outcome, and Fig. 12b shows the ALL-IDB2 image result. These results confirm that the VGG-UNet helps to get a comparatively better outcome on both the datasets.

Fig. 11
figure 11

Results achieved with BCCD and ALL-IDB2

Full size image
Fig. 12
figure 12

Comparison of average performance achieved for 250 images

Full size image

The future scope of our research work may be concentrated toward improving the performance of the CNN schemes by adjusting the initial parameters, such as adjusting the image augmentation process, improving the learning, modifying the decoder-encoder batch, adjusting the weight initialization process, modifying the dropout rate, and modifying the activation layer.

This work employed encoder-decoder (VGG-UNet) scheme to achieve accurate segmentation of the leukocyte region in RGB-scaled image. The outcome of the encoder section will be the learned features (Deep-Features), and this feature can be considered to classify the images using a binary or multiclass classifiers. The future scope of this research includes: (1) Implementation of VGG supported automated image classification, (2) Development of VGG-SegNet, and (3) Examination of clinical-grade images.

5 Conclusion

Due to its medical importance, a significant amount of image assessment schemes is planned and implemented by the researches for the medical images with varied modalities. This research presented and automated leukocyte extraction system from hematological images using the benchmark LISC dataset. This work used the pre-trained CNN segmentation procedures, such as SegNet, U-Net, and VGG-UNet to extort the leukocyte section with better accuracy. The proposed segmentation procedure is implemented on 250 images associated with the GT and before implementing the segmentation process, every test image is resized into \(256x256x3\) pixels, in order to reduce the computation burden. The results are then compared to identify which CNN scheme produced the best outcome for the LISC dataset. Our experimental results show that the VGG-UNet produced better results than SegNet and U-Net. The outcome of the VGG-UNet is also authenticated against the other hybrid segmentation procedures existing in the literature. Further, the eminence of proposed scheme is tested and validated on BCCD and ALL-IDB2 and these results also verify that this technique helps to segment the leukocyte image perfectly. This research work demonstrated that CNN is useful and is significant in the clinical domain. In future, we will examine the clinical grade of hematological images. Furthermore, the proposed VGG-UNet approach can also be considered to classify the considered image database with a binary and multiclass classifier to support the automated leukocyte class recognition.