Acute lymphoblastic leukemia cancer diagnosis in children and adults using transforming blood fluorescence microscopy imaging

Dataset description

This work provides evidence of the development of ALL, a data collection method that is both comprehensive and accessible to the general public. We utilized the previously described dataset for training and validate the proposed methodology. The bone marrow laboratory at Taleqani Hospital, located in Tehran, Iran, created the images included in this collection. There are 3,256 PBS photographs in the collection. With meticulous attention to detail, laboratory personnel took these photographs of dyed and processed blood samples (Aria et al., 2021). They collected samples from 89 individuals suspected of having acute lymphoblastic leukemia. Hematogones make up the first group, while early, benign cells, pre-cells, and pro-cells make up the subsequent groups. It is common practice to refer to all of the three unique subtypes of malignant lymphoblasts that make up the AL group as “ALL.” They linked a Zeiss camera to a microscope with a hundred-fold magnification to capture the images. After that, the file system saved the images as JPG files. A specialist used the flow cytometry apparatus to gain a definitive assessment of the cell types and subtypes. Ultimately, deliver segmented images that undergo color thresholding in the HSV color space, resulting in their separation into distinct segments. Figure 2 displays a few of samples from the acquired dataset.

Image dataset preprocessing

The most efficient and practical strategy for enhancing the efficiency and practicality of image preparation in experiments was found to be cutting, as opposed to scaling. Edges of an image, which are sudden disruptions, may contain almost as much information as the pixels themselves. Preprocessing is very important in deep learning, especially for medical imaging, to clean the data and make it suitable for further processing. The image datasets have a large number of dimensions. When we input them into the model, the model takes too much training time, so to minimise the processing time and reduce the computational cost, we first resize the data into smaller dimensions. This hinders the ability to discern the most prominent characteristics. An image pixel’s hold information that changes when its size increases. The approximation process enlarges an image to create a large-scale image. The utilised photographs have a size of 224 $\times$ 224 pixels. Processing such enormous photos is simply impractical due to the substantial CPU requirements associated with increasing the pixel count of an image (Omar Bappi et al., 2024).

Image dataset augmentation

Deep networks in deep learning need a large amount of training data if they are to properly generalize and achieve outstanding accuracy. Sometimes, nevertheless, the scale of image data is inadequate. The image augmentation techniques are shown in Algorithm 1. The system generates synthetic training data by applying various techniques such as zooming, flipping, random rotation, and contrast modification to the provided data. Improving the performance of deep learning algorithms depends on a more extensive dataset with a complex network architecture, as the efficacy of the training dataset greatly influences their performance. To accommodate the common constraints seen in medical imaging datasets, researchers use rotation, shifting, flipping, and cropping among various data augmentation methods. Researchers define cropping in terms of patches by removing extra material and using rotation angles. we also used certain frameworks employ flipping methods and shifting operations to enhance the training dataset (Koshinuma et al., 2023). Figure 3 displays the class distribution of training data with and without augmentation.

ResNet-CNN architecture

AlexNet had the distinction of being the first CNN architecture to emerge victorious in the ImageNet 2012 competition. As a result, each successful model has included more layers in a deep neural network with the goal of reducing the error rate. This approach is effective when there are a smaller number of levels, but it has difficulties with the disappearing or exploding gradient when there are a larger number of layers. The field of deep learning frequently confronts this issue. The gradient either diminishes or significantly magnifies. As the number of layers increases, there is a corresponding increase in the error rates during both training and testing. The proposed methodology for ALL diagnosis is presented in Algorithm 2. In this architecture, the use of residual blocks effectively resolved the issue of slopes either diminishing or amplifying. In the network layer hierarchy, certain levels between the link layer and the subsequent layers do not activate the skip connection. Consequently, a residual block is created. The residual network, often known as ResNet, plays a crucial role in addressing issues related to computer vision. The ImageNet dataset, comprising 1,000 distinct categories of objects, served as the training dataset for this network. This effectively demonstrated how the remaining blocks handled the provided photographs. Every block has several levels. A layer-by-layer description of the ResNet-CNN architecture as well as its parameters are shown in Table 2.

Figure 4 depicts the architectural plan for the structure’s development. The dimensions of the input are 224 by 224, and it consists of three color channels: red, green, and blue. The framework included the top layer, ResNet, in the ResNet set to facilitate the creation of a sequence model. It was crucial since we relied on our own customized levels determined by the datasets. Employing ResNet and imageNet weights in conjunction simplifies the process of transfer learning. Layers such as convolutional, pooling, activation, and fully connected will be used for testing and training purposes. Subsequently, the CNN model employs three conv2D layers, one Maxpool layer, and three batch normalization layers. Next, we include several specific layers into the framework, such as a global pooling average layer, batch normalization layer, relu function dense layer, dropout layer, two more relu function dense layers, another dropout layer, and ultimately a softmax multiclass classification dense layer.

The convolutional layer is a crucial component of any convolutional neural network. The CONV layer consists of K-trainable filters, commonly known as “kernels,” which may have square dimensions for both width and height. Regardless of their size, these displays completely encompass the entire quantity. In RGB photographs, the depth of the data indicates the number of channels present in the image. The number of filters applied in the preceding layer determines the volume levels in subsequent layers of the network. Batch normalization is feasible because the normalization technique calculates the means and variances of each input layer. The entire training set will be perfectly suitable. Dropout is a regularization technique that aims to prevent models from overfitting by improving accuracy while potentially compromising training accuracy. Dropout layers in our training set stochastically eliminate inputs from one layer to the next in the network architecture for each mini-batch, with a Prob of p. A typical practice is to place the pooling layer behind the convolution layer in order to decrease the spatial dimension of the data. We process the components at different depths of the input volume independently. Pooling techniques securely maintain the volume depth.

Model selection

In this study, we select ResNet-CNN because of its excellent performance in image classification, especially in medical cancer classification. ResNet uses special connections that remove noise during training, which allows us to build deep networks without compromising performance. Compared to traditional CNNs or other architectures such as VGG-16 or EfficientNet, ResNet-CNN offers excellent depth, computational speed, and accuracy. Its ability to learn useful features from high-quality medical images makes it a prime fit for the early detection of ALL in blood images. We apply batch normalization layers to stabilize the training, followed by a global pool to effectively reduce the size of the feature mapping space. Dropout layers are added to prevent overfitting. In addition, its design facilitates data accessibility and automation, which is important when working with medical datasets of varying size and quality.

Deep classifiers

MobileNetV1 achieved parameter reduction by using a depth-wise convolution. To implement a system that incorporates three levels of filtering, compression, and expansion, an expansion layer was included during the second block iteration. To further enhance performance optimization, the Inverted Residual Block technique was used. Convolutional neural networks such as Mobilenet are highly suitable for mobile vision applications due to their user-friendly nature, low computational requirements, and ability to provide satisfactory outcomes. MobileNet is used in several practical applications, including geolocation, face analysis, detailed categorization, and item identification. Depth-wise convolutions are used to apply a single filter to each input channel (Mondal et al., 2021). During a typical convolution, the filters are applied separately to each input channel.

The visual geometry group (VGG) used six different neurons; Results were obtained for the VGG16 (Sriram et al., 2022) and VGG19 (Ahmed & Nayak, 2021) models. Three 3 $\times$ 3 blocks of size 7 $\times$ 7 were used, with two blocks in a row, to give a more predictable 5 $\times$ 5 pattern. A maximum pooling operation yields differences in regression coefficients. The VGG16 network is composed of two parallel and two parallel layers. In addition, VGG19 has sixteen layers and three fully connected layers. Entire VGG networks use 3–3 layers with 1 linear layer and several non-linear layers. VGG-19 shows that in order to achieve the highest possible classification accuracy, eighteen deep layers are required.

The fundamental concept behind EfficientNet’s convolutional neural network is known as “compound scaling.” This concept ultimately resolves the long-standing dilemma of how to maximize the efficiency of model dimensions, computational time, and precision. Compound scaling enables the modification of the depth, resolution, and breadth of a neural network, which are three crucial attributes of the network (Das, Pradhan & Meher, 2021). The breadth of a neural network refers to the amount of channels present in a certain layer. Increasing the breadth of the model enhances its accuracy by enabling it to capture a greater number of complex patterns and features. Conversely, the model’s decreased bulk enhances its practicality in locations with limited resources. When it comes to increasing the depth of a network, the number of network layers is of utmost importance. Despite requiring more computational resources, more intricate models have the capability to collect more intricate data pictures. Although more complex models may need fewer computational resources, their level of accuracy may be lower. Resolution scaling refers to the process of adjusting the size of an image to a new dimension. Images with a higher resolution exhibit greater levels of detail, perhaps resulting in enhanced performance. However, they need more RAM and computational capacity. Conversely, lower-quality images are more economical in terms of resources, but they may result in the loss of delicate details.

Google researchers have developed an InceptionV3 architecture specifically designed for CNN. In 2015, the InceptionV3 architecture was introduced as the successor to InceptionV1 and InceptionV2. InceptionV3 is specifically designed for image classification, offering a combination of high computational efficiency and unparalleled accuracy (Ramaneswaran et al., 2021). Image feature extraction is easy using the InceptionV3 architecture’s full suite of transform, fusion, and initialization modules. The network can capture a series of different features at different scales and resolutions by using the initialization module, including features by using filters of different sizes. In the domains of visual query answer, object detection, and image classification, the performance of InceptionV3 consistently outperforms existing advanced models. Transfer learning is a technique that improves performance on a specific task by adapting pre-trained weights to a new data set. Typically, it uses the prior trend weighting as the central model.

Performance evaluation

Performance evaluations are crucial for assessing the efficacy and utility of deep learning classifiers. They are important in the learning process because they help programmers improve their systems and progress further. This may be difficult to choose, but choosing a fair and accurate performance metric for your work is crucial. Classification metrics evaluate the learning models’ performance in a variety of classification tasks. Their objective is to allocate each individual data point to one or more predetermined categories. An essential criterion for evaluating performance of deep model for detection tasks is its accuracy. The proportion of accurately predicted occurrences remains constant in relation to the total number of events in the dataset. Precision is defined as the ratio of correct positive estimates to the total number of positive forecasts (Rismayanti et al., 2023). The recall rate is defined as the proportion of positive predictions out of all actual positive instances. This test evaluates the classifier’s ability to accurately identify high-quality instances. A model with high accuracy demonstrates a reduced occurrence of both false_positives and false_negatives. One of these indicators may be more important, depending on the situation. The F1-score is a metric that maintains a balance between precision and recall by calculating the harmonic mean of the two. When one class is much more prevalent than the other, working with imbalanced data may be advantageous.

(1) $$Accuracy=\frac{(TRPR+FANR)}{(TRRP+TRNR+FAPR+FANR)}$$

(2) $$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}=\frac{TRPR}{(TRPR+FAPR)}$$

(3) $$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}{\mathrm{l}}_{\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}=\frac{TRPR}{(TRPR+FANR)}$$

(4) $$\mathrm{F}{1}_{\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}=(2\times \frac{{\displaystyle (\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}})\times ({\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}_{\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}})}}{(\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}})+(\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}{\mathrm{l}}_{\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}})}).$$

Acute lymphoblastic leukemia diagnosis analysis

The classification report for the proposed method for the detection of ALL and analysis is provided in Table 4. This study demonstrates the performance of four sub-types of leukemia cancer using an original public dataset of microscopy blood samples. The pro- cancer subtype exhibited outstanding performance, scoring 100% in all settings. Additionally, the early cancer subtype obtained $precisio{n}_{score}$, $recal{l}_{score}$, and $f{1}_{score}$ of 96%. The benign cancer earned a recall rate of 99% and a F1-score of 96%. The pre-cancer detection model attained an accuracy rate of 100% and an $f{1}_{score}$ of 98.42%. Nevertheless, the benign cancer has a low performance compared to other cancer subtypes, with a precision of just 93%. The proposed method achieved an overall accuracy of 97.85% and a macro average $f{1}_{score}$ of 98.08%.

Table 5 presents the categorization report for the proposed method to acute lymphoblastic leukemia detection and analysis. This research uses an improved dataset of microscopy blood samples to illustrate the characteristics of four subtypes of leukemia malignancy. The pro-cancer subtype performed very well, achieving $recal{l}_{score}$ of 100%, $precisio{n}_{scores}$ of 99.68%, and $f{1}_{score}$ of 99.84%. Furthermore, the $precisio{n}_{score}$, $recal{l}_{score}$, and $f{1}_{score}$ for the early cancer subtype were 97.40%, 98.16%, and 97.78%, respectively. The benign cancer received $f{1}_{score}$ of 97.82% and a $recal{l}_{rate}$ of 97.03%. With an $f{1}_{score}$ of 99.46% and an $precisio{n}_{score}$ of 99.19%, the pre-cancer detection model achieved success. The proposed technique performed well, with a macro average $recal{l}_{score}$ of 98.73% and an overall accuracy of 98.67%.

Table 6 calculates the results and variability of the last epochs. The highest level of precision is 98.77, achieved during epoch 9. Over the last 10 epochs, the average accuracy has ranged from 98.77 to 0.405 standard deviations.

Result analysis using accuracy and loss

Evaluating the accuracy learning_curves and loss_curves of the proposed approach are crucial in order to validate the efficacy of the detection findings in imaging field and healthcare. Figure 5 represents the epoch wise accuracy and loss during training and validation for the proposed technique. The loss, whether during training or validation, provide some sort of indicator of performance affects and provides information on the error rate. Figure 5A demonstrates that the proposed approach using the original data has worse performance compared to the augmented dataset shown in Fig. 5B. Figure 5C displays the loss while using the original dataset, whereas Fig. 5D exhibits the loss when using an augmented dataset of blood microscopy. During the first epochs, the validation loss exhibited a high value, but subsequently decreased and eventually dropped below 0.07.

The receiver operating characteristic (ROC) curve typically displays TPR on Y_axis and FPR on the X_axis. The top left_corner of Fig. 6 defines the best location with a TPR of one and a FPR of zero. Consequently, having a bigger AUC is sometimes more desirable, despite the fact that it is not practicable to do so. Because the optimal strategy involves maximizing the TPR and decreasing the FPR, the steepness of ROC curves becomes an essential factor. Because the TPR and the FPR are so well-defined and easy to understand, ROC curves are often used in classification. Figure 6A displays ROC-AUC using the original data, whereas Fig. 6B shows the ROC-AUC using the enhanced data. The performance was boosted by using additional data.

Precision-recall curves of the proposed technique with and without augmentation is shown in Fig. 7. The precision and recall scores may vary depending on the criteria. The threshold is the lowest prob required for classifying a prediction as positive. Enhancing the thresholds often leads to enhanced precision, but it often causes a decrease in the model’s recall.

Cumulative gain is a numerical metric used to ensure the efficiency of a model in the domains of machine learning and information retrieval, particularly in tasks involving ranking is shown in Fig. 8. This approach simplifies the evaluation of the model’s ability to rank things accurately by combining the gains (or relevance scores) of items from several metrics.

A confusion matrix, also known as an error matrix, is a specific tabular arrangement that provides a visual representation of the outcomes of an algorithm, often used in supervised learning (unsupervised learning uses a distinct term, such as matching matrices). Results of normalized-confusion matrix with and without augmentation is shown in Fig. 9. The matrix contains actual classes in its rows and predicted classes in its columns.

Grad-CAM

Grad-CAM is a visual tool that presents choices made using a specified framework in a clear and concise manner. To predict classes in advance, it generates course localization maps that emphasize significant sections of the input image. Grad-CAM significantly streamlines the task of seeing and comprehending the understanding process of deep transfer learning models. Grad-CAM is a technique that identifies the specific areas in an image that impact the model’s prediction. This helps to improve the understanding, clarity, and reliability of the model’s decision-making process. This method offers more flexibility and precision compared to earlier procedures. Fortunately, despite its intricate nature, the outcome is unambiguous. To construct a Grad-CAM heat map, we generate a model by segmenting a high-level picture at the desired layer in order to construct a Grad-CAM heat map. We provide the entire set of interconnected layers in order to make predictions. Subsequently, we collect the output_layer, implement the loss_function, and input it into the model. We derive the gradient of our proposed model layer with regard to the model loss. Next, to superimpose the heat map onto the original images, we remove the areas that support the gradient, decrease their size, resize them, and adjust their scale. The Grad-CAM are visualized in Figs. 10A and 10B.

Analysis of transfer learning pretrained classifiers is crucial in classification and detection. Deep classifiers apply a method known as transfer learning to solve one problem using an already-trained model for another. By applying what it has learned in one task to another, the classifier may improve its generalizability and overall performance via a process known as transfer learning. The data gathered may be used to detect cancer while training a classifier to predict whether an image contains tumors or not. Generally, a neural network should dedicate its first few layers to edge detection, the middle layer to shape recognition, and the final layer to task-specific feature extraction. The process of transfer learning entails making use of the first and intermediate layers while retraining just the subsequent ones. It helps to make the most of the labeled data from the first training session. The retraining process refines the models. However, in order to retrain using transfer learning, one must isolate and concentrate on certain layers. Table 7 displays the experiments on deep classifiers.

Comparison with existing studies

Deep transfer learning and fusion approaches provide a significant improvement compared to their previous versions. Previously, classical machine learning techniques like random forest and manually chosen feature sets were used. However, their use was limited by the non-uniform nature of anatomic-pathological data sets and their reliance on domain expertise. In addition, many algorithms need several pre-processing and feature extraction phases, which may be a lengthy procedure. Moreover, mistakes might arise due to human intervention. Comparison with existing studies is shown in Table 8. Genovese et al. (2021) used 260 blood samples from ALL databases and obtained an accuracy of 96.8% using VGG-16. However, the attained accuracy is considered to be somewhat poor. Das & Meher (2021), also obtained a very low level of accuracy while using the random forest methodology. Khan Tusar et al. (2024) used DNN and obtained a 97% accuracy rate. Bhuvaneswari et al. (2023) used the VGGNet model and obtained a remarkable accuracy of 95.7%. Najjar et al. (2023) obtained a 96.1% accuracy by using the hue saturation value method. Keerthivasan & Saranya (2023) used a modified CNN to diagnose cancer and obtained a remarkable accuracy of 97.6% by using all 3256 samples. Dangore et al. (2024) used CNN and attained a remarkable accuracy of 96.1%. Visual representation of comparison with existing studies is shown in Fig. 11. The previous research used same datasets and samples to diagnose ALL cancer, although yielded less precise results. The proposed strategy produced superior accuracy by using the proposed preprocessing and sophisticated augmentation approaches, as deep learning demonstrates enhanced performance with larger datasets.

Statistical T-test

We used a statistical t-test to measure the difference in the performance of the proposed model using augmented and without-augmented data as illustrated in Table 9. In all cases, we utilised the last five epochs’ accuracy, precision, and recall of the proposed model with and without augmentation and applied a t-test. We obtained t-test values of 3.97004, 3.43585, and 3.46484, along with p-values of 0.01654, 0.02639, and 0.02570 for accuracy, precision, and recall, respectively. The test indicates that if the p-value is <0.05, there is a significant difference, while if the p-value is >0.05, then there is no difference. In our case, our values indicate that there is a significant difference in the performance when comparing with the augmented data and improvement in the accuracy. Also, in case 1, the first row indicates the mean score + STD without augmentation, and the 2nd row indicates with augmentation. Also, using the mean score, we obtained 0.0062 STD without augmentation and 0.0039 STD with augmentation, ensuring that with augmentation, the proposed approach achieved a minimum STD score.