In recent years, huge volumes of healthcare data are getting generated in various forms. The advancements made in medical imaging are tremendous owing to which biomedical image acquisition has become easier and quicker. Due to such massive generation of big data, the utilization of new methods based on Big Data Analytics (BDA), Machine Learning (ML), and Artificial Intelligence (AI) have become essential. In this aspect, the current research work develops a new Big Data Analytics with Cat Swarm Optimization based deep Learning (BDA-CSODL) technique for medical image classification on Apache Spark environment. The aim of the proposed BDA-CSODL technique is to classify the medical images and diagnose the disease accurately. BDA-CSODL technique involves different stages of operations such as preprocessing, segmentation, feature extraction, and classification. In addition, BDA-CSODL technique also follows multi-level thresholding-based image segmentation approach for the detection of infected regions in medical image. Moreover, a deep convolutional neural network-based Inception v3 method is utilized in this study as feature extractor. Stochastic Gradient Descent (SGD) model is used for parameter tuning process. Furthermore, CSO with Long Short-Term Memory (CSO-LSTM) model is employed as a classification model to determine the appropriate class labels to it. Both SGD and CSO design approaches help in improving the overall image classification performance of the proposed BDA-CSODL technique. A wide range of simulations was conducted on benchmark medical image datasets and the comprehensive comparative results demonstrate the supremacy of the proposed BDA-CSODL technique under different measures.
Big data originally demonstrates the variety, volume, and velocity of data acquired during different data production times. Medical big data, acquired at healthcare providers, contains data relevant to patient care such as diagnoses, demographics, medications, medical procedures, immunizations, vital signs, radiology images, laboratory results and so on [
The huge amount of health care data constitute the images acquired through medical imaging techniques (Echography, Computed Tomography Scan, MRI, Mammography, and so on.). In order to perform a comprehensive analysis and achieve biomedical image management, each step in healthcare process should be automated [
ML and DL methods (i.e., CNN, SVM, and NN) have attained outstanding performances in biomedical image classification [
In this background, the current study presents a new Big Data Analytics with Cat Swarm Optimization based Deep Learning (BDA-CSODL) technique for medical image classification on Apache Spark environment. The proposed BDA-CSODL technique involves Bilateral Filtering (BF)-based noise removal technique as a preprocessing step. Moreover, the proposed BDA-CSODL technique involves multi-level thresholding-based image segmentation too with deep convolutional neural network-based Inception v3 technique as a feature extractor. Furthermore, Stochastic Gradient Descent (SGD) model is used for parameter tuning process. Finally, CSO with Long Short-Term Memory (CSO-LSTM) method is utilized as a classifier to allot proper class labels to it. In order to showcase the improved performance of the proposed BDA-CSODL technique, a comprehensive experimental analysis was conducted against benchmark medical images.
In Ashraf et al. [
In Tan et al. [
Wang et al. [
In this study, a new BDA-CSODL technique is presented for biomedical image classification on Apache Spark environment. The proposed BDA-CSODL technique encompasses BF-based preprocessing, Otsu-based segmentation, Inception v3-based feature extraction, SGD-based hyperparameter optimization, LSTM-based classification, and CSO-based parameter tuning.
Apache Spark is a distributed computing system utilized in big data environment and is the most effective framework developed in this regard. Spark provides a complete and unified structure for managing differential needs to big data processing with different types of datasets (text data, graph data, image/video, and so on.) collected from various sources (realtime streaming, batch, etc.). Spark has been developed to resolve the problem encountered in Hadoop frameworks. Indeed, Spark frameworks have demonstrated its fast-execution potentials than Hadoop frameworks under different scenarios (over 100 times in memory).
Apache Spark framework involves several phases as shown in
Since the number of features could be dissimilar based on the image, some clauses can be included so as to make these feature vectors are of similar size. Next, vector descriptors are constructed according to this feature; all the descriptors are of similar size. In this method, it has to be pointed out that the feature extraction from labeled/unlabeled images is executed by several images in big data context, regarding distinct V's of big data (velocity, volume, variability, veracity, and variety). But the performance of Spark algorithm gets reduced at few scenarios, particularly during feature extraction when few smaller images exist under the datasets (labeled biomedical images or unlabeled biomedical image). In order to resolve this problem, the researchers proposed two approaches such as feature extraction by segmentation and sequence in feature extraction. The performance of one of these two approaches could solve the problems faced in unbalanced loading, while the execution time of every task could be a similar process. The classification procedure is initiated, once the novel unlabeled data comes to the scheme.
In this method, both query and prediction are implemented during MapReduce stage itself. The adoption of Spark frameworks comes with advantage to process in a big data platform and uses their embedding libraries, for instance MLlib (Machine Learning library).
For a non-linear, edge preserving image filtering model, BLF treats the intensity values of all the pixels as weighted average of its adjacent pixels’ intensity value [
Here,
In
Next to preprocessing, Otsu technique is employed to segment the medical images and determine the affected regions. Otsu is a generally-employed image thresholding method that was originally proposed by Otsu in 1979. At present, Otsu criterion is commonly used in defining the optimum threshold model which can offer histogram-based image segmentation using acceptable results. For simple bi-level threshold, the original method in Otsu can be determined as following: initially, an image is assumed to have
When the threshold is
The mean intensity values for bi-level threshold of all the classes are shown in the following equation
After calculating this value, class variance
Whereas
Whereas
Whereas 1 ≤
During feature extraction, Inception v3 model is executed to produce a useful set of feature vectors. In line with GoogLeNet (Inception-v1), Inception-v3 has progressed well in recent years with each update being incorporated in Inception-v2. During classification, LSR (Label Smoothing Regularization) is included after fully-connected layers. In convolution layer, 7 × 7 kernels are substituted by 3 × 3 kernels. Also, regularization is utilized and normalization is included in loss function to avoid overfitting. Initially, three convolution layers, two pooling layers, one pooling layer and two convolution layers are fixed. At last, elven dropout layer, mixed layers, softmax layer, and fully connected layer are followed. Padding and convolution operations are performed continuously on the images, in all the layers [
Since hyperparameter tuning process affects the classification performance considerably, it is performed using SGD optimizer. With reference to determination of mean variance portfolio optimization problems, an augmented objective function
Whereas
Note that the first order necessary condition for
The step size
In final stage, image classification process is perfumed using CSO-LSTM model. Traditional neural networks such as CNN perform better in extracting invariant feature. However, when it comes to prediction of present output condition on long distance features, RNN performs well than CNN. In each RNN unit, the input represents the feature at time t while the dimension is denoted as feature size. Hidden state ℎ−T represents the memory beforehand in this unit. With ℎ−T and input, one can compute and transfer the novel memory ℎ to the succeeding RNN unit. Thus, RNN has a problem in exploiting and finding long term dependency in the dataset. While handling sequential data, RNN occasionally meets challenges such as gradient vanishment and gradient explosion. LSTM is a technique designed from RNN that produce outstanding performances when handling long term memory. As displayed,
The matrix, in the previous equation, has the same meaning as its name. E.g., Who implies hidden input gate matrix. After processing in LSTM unit, it passes the hidden state ℎt to the following LSTM unit. It is accountable for resolving the signal of upcoming time slice and sending its output to the following layer. Since LSTM continues to have certain units such as forget gate, this technique could remember significant long-term memory while it can also adopt to short-term memory as well that has significant data.
In order to determine the parameters involved in LSTM model optimally, CSO algorithm is used. In spite of spending most of its time taking rest, cats exhibit high curiosity and awareness regarding moving objects and surroundings in their environments. Such behaviors help cats in locating prey and hunt them down. In comparison with the time devoted to their resting, cats use minimal time to chase their prey and save their energy. Stimulated by these hunting patterns, Chu et al. [
Next, the computation procedure of CSO is described in a step-wise fashion herewith.
Step1. Generate the first population of cats and separate them into
Step2. Based on the value of
Step3. Compute the fitness value of all the cats and save the cat with optimal FF. The location of the optimal cat (
Step4. According to their flag, the cats are to be employed under tracing/seeking modes as follows.
Step5. Once the end conditions are fulfilled, end the procedure. Or else repeat the steps 2 via 5.
In seeking mode, the cats tend to take rest, while it also keeps a track of their surroundings. When it senses a prey or a danger, cats decide their next move. Similar to resting, in this seeking mode, the cat observes
Next, the seeking mode is determined as follows.
Step 1: Create SMP copy of every
Step 2: For all the copies, evaluate a novel location via
Step 3: Calculate the Fitness Value (
Step 4: With roulette wheel, the points are arbitrarily selected to move towards the candidate point and the location of
This mode simulates how cat chases its prey. Afterward detecting a prey when resting (seeking), the cat decides their motion direction and speed according to the prey's speed and position. In CSO, the velocity of cat
The termination criteria is defined, once the process is ended. The selection of an end condition plays a significant role in guaranteeing accurate convergence. The number of iterations, number of development, and the execution time are general end conditions for CSO algorithm.
This section details about the results achieved from experimental validation of the proposed BDA-CSODL technique on applied CT brain, chest, and cervical images. For simulation, a set of 10,000 images under every dataset was applied. The results were investigated in terms of two measures namely, kappa and accuracy.
Accuracy (%) | |||
---|---|---|---|
Sample images | Brain | Chest | Cervical |
2000 | 70.01 | 63.84 | 80.01 |
4000 | 88.48 | 79.51 | 83.24 |
6000 | 89.22 | 87.64 | 89.27 |
8000 | 95.92 | 92.99 | 97.48 |
10000 | 99.78 | 94.49 | 98.96 |
Kappa (%) | |||
---|---|---|---|
Sample number | Brain | Chest | Cervical |
2000 | 64.87 | 63.81 | 74.71 |
4000 | 87.75 | 75.99 | 83.10 |
6000 | 90.08 | 85.62 | 87.86 |
8000 | 95.57 | 91.11 | 96.65 |
10000 | 99.67 | 92.84 | 98.85 |
A brief comparison study was conducted between BDA-CSODL technique against existing approaches on brain dataset and the results are shown in
Methods | I-2000 | I-4000 | I-6000 | I-8000 | I-10000 |
---|---|---|---|---|---|
Accuracy | |||||
BDA-CSODL | 0.700 | 0.885 | 0.892 | 0.959 | 0.998 |
DL-GAN | 0.695 | 0.879 | 0.889 | 0.954 | 0.996 |
DL-CNN | 0.585 | 0.828 | 0.788 | 0.953 | 0.988 |
FR-CNN | 0.547 | 0.880 | 0.863 | 0.937 | 0.981 |
CNNSVM | 0.535 | 0.800 | 0.858 | 0.885 | 0.884 |
SVM | 0.562 | 0.798 | 0.820 | 0.945 | 0.975 |
Kappa | |||||
BDA-CSODL | 0.649 | 0.878 | 0.901 | 0.956 | 0.997 |
DL-GAN | 0.644 | 0.853 | 0.898 | 0.953 | 0.995 |
DL-CNN | 0.579 | 0.808 | 0.829 | 0.951 | 0.984 |
FR-CNN | 0.563 | 0.873 | 0.876 | 0.933 | 0.980 |
CNNSVM | 0.526 | 0.779 | 0.844 | 0.875 | 0.868 |
SVM | 0.568 | 0.796 | 0.811 | 0.941 | 0.975 |
With respect to kappa, the proposed BDA-CSODL algorithm accomplished the maximum performance, whereas CNNSVM, FR-CNN, and SVM methods gained the least performance. For instance, for I-2000 samples, BDA-CSODL method accomplished effectual results with a maximum kappa of 0.649, whereas DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM methods achieved low kappa values such as 0.644, 0.579, 0.563, 0.526, and 0.568 respectively. In the meantime, with I-6000 samples, BDA-CSODL system accomplished an effective outcome with a high kappa value i.e., 0.901. However, DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM approaches gained minimal kappa values namely, 0.898, 0.829, 0.876, 0.844, and 0.811. Eventually, with I-10000 samples, the proposed BDA-CSODL method accomplished an efficient outcome with high kappa of 0.997, whereas other techniques such as DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM methods achieved minimum kappa values namely, 0.995, 0.984, 0.980, 0.868, and 0.975 correspondingly.
A detailed comparative analysis was conducted between BDA-CSODL approach and existing methods on chest dataset and the results are shown in
Methods | I-2000 | I-4000 | I-6000 | I-8000 | I-10000 |
---|---|---|---|---|---|
Accuracy | |||||
BDA-CSODL | 0.638 | 0.795 | 0.876 | 0.930 | 0.945 |
DL-GAN | 0.634 | 0.754 | 0.868 | 0.927 | 0.941 |
DL-CNN | 0.615 | 0.775 | 0.857 | 0.894 | 0.908 |
FR-CNN | 0.589 | 0.789 | 0.863 | 0.905 | 0.917 |
CNNSVM | 0.437 | 0.665 | 0.765 | 0.811 | 0.894 |
SVM | 0.597 | 0.792 | 0.872 | 0.924 | 0.922 |
Kappa | |||||
BDA-CSODL | 0.638 | 0.760 | 0.856 | 0.911 | 0.928 |
DL-GAN | 0.633 | 0.744 | 0.843 | 0.908 | 0.924 |
DL-CNN | 0.629 | 0.739 | 0.847 | 0.897 | 0.887 |
FR-CNN | 0.468 | 0.754 | 0.837 | 0.892 | 0.895 |
CNNSVM | 0.516 | 0.590 | 0.724 | 0.814 | 0.883 |
SVM | 0.475 | 0.755 | 0.853 | 0.905 | 0.907 |
In terms of accuracy, the proposed BDA-CSODL method accomplished a high efficiency whereas CNNSVM, FR-CNN, and SVM methods gained low performance. For instance, with I-2000 samples, the proposed BDA-CSODL algorithm accomplished effectual results with a maximum accuracy of 0.638, whereas other techniques such as DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM gained minimal accuracy values such as 0.634, 0.615, 0.589, 0.437, and 0.597 correspondingly. In addition, with I-6000 samples, BDA-CSODL algorithm accomplished effectual results with a high accuracy of 0.876, whereas DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM systems obtained the least accuracy values such as 0.868, 0.857, 0.863, 0.765, and 0.872 correspondingly. Also, with I-10000 samples, the proposed BDA-CSODL technique accomplished an effectual outcome with a high accuracy of 0.945, whereas DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM algorithms gained minimum accuracy values namely, 0.941, 0.908, 0.917, 0.894, and 0.922.
With respect to kappa, BDA-CSODL approach accomplished the maximal performance whereas other techniques such as CNNSVM, FR-CNN, and SVM achieved minimal performance. For instance, with I-2000 samples, the proposed BDA-CSODL system accomplished effective outcomes with an improved kappa of 0.638, whereas DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM methods attained minimum kappa values namely, 0.633, 0.629, 0.468, 0.516, and 0.475 correspondingly. At the same time, with I-6000 samples, the proposed BDA-CSODL approach accomplished efficient outcomes with a high kappa of 0.856, whereas DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM methods gained minimal kappa values namely, 0.843, 0.847, 0.837, 0.724, and 0.853. Finally, with I-10000 samples, the presented BDA-CSODL methodology accomplished efficient results with a high kappa of 0.928, whereas other techniques such as DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM algorithms gained minimum kappa values such as 0.924, 0.887, 0.895, 0.883, and 0.907 correspondingly.
A brief comparative analysis was conducted between the proposed BDA-CSODL approach against recent algorithms on cervical dataset and the results are shown in
Methods | I-2000 | I-4000 | I-6000 | I-8000 | I-10000 |
---|---|---|---|---|---|
Accuracy | |||||
BDA-CSODL | 0.800 | 0.832 | 0.893 | 0.975 | 0.990 |
DL-GAN | 0.796 | 0.828 | 0.890 | 0.970 | 0.985 |
DL-CNN | 0.656 | 0.778 | 0.858 | 0.958 | 0.976 |
FR-CNN | 0.698 | 0.789 | 0.866 | 0.964 | 0.983 |
CNNSVM | 0.624 | 0.688 | 0.763 | 0.906 | 0.974 |
SVM | 0.766 | 0.825 | 0.888 | 0.963 | 0.974 |
Kappa | |||||
BDA-CSODL | 0.747 | 0.831 | 0.879 | 0.967 | 0.989 |
DL-GAN | 0.741 | 0.812 | 0.875 | 0.962 | 0.984 |
DL-CNN | 0.683 | 0.682 | 0.844 | 0.950 | 0.975 |
FR-CNN | 0.662 | 0.747 | 0.849 | 0.957 | 0.982 |
CNNSVM | 0.616 | 0.648 | 0.745 | 0.894 | 0.968 |
SVM | 0.743 | 0.827 | 0.873 | 0.961 | 0.983 |
In terms of accuracy, the proposed BDA-CSODL methodology accomplished a maximal performance outcome whereas CNNSVM, FR-CNN, and SVM approaches gained minimal performance outcomes. For instance, with I-2000 samples, BDA-CSODL methodology accomplished an effective outcome with an increased accuracy of 0.800, whereas other techniques such as DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM methods gained minimal accuracy values such as 0.796, 0.656, 0.698, 0.624, and 0.766 correspondingly. Along with that, for I-6000 samples, the proposed BDA-CSODL technique accomplished effective results with enhanced accuracy i.e., 0.893, whereas other techniques such as DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM approaches achieved the least accuracy values such as 0.890, 0.858, 0.866, 0.763, and 0.888 correspondingly. Meanwhile, with I-10000 samples, the proposed BDA-CSODL system accomplished efficient outcomes with a superior accuracy of 0.990, whereas DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM techniques attained minimum accuracy values namely, 0.985, 0.976, 0.983, 0.974, and 0.974.
In terms of kappa, the proposed BDA-CSODL approach accomplished a high efficiency, whereas CNNSVM, FR-CNN, and SVM methodologies gained the least efficiency values. For instance, with I-2000 samples, BDA-CSODL method accomplished effective results with an increased kappa of 0.747, whereas DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM systems achieved minimum kappa such as 0.741, 0.683, 0.662, 0.616, and 0.743 correspondingly. Simultaneously, with I-6000 samples, the proposed BDA-CSODL approach accomplished an efficient outcome with a maximum kappa of 0.879, whereas DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM techniques gained low kappa values namely, 0.875, 0.844, 0.849, 0.745, and 0.873. Concurrently, with I-10000 samples, the proposed BDA-CSODL method accomplished effectual outcomes with a superior kappa of 0.989. However, other techniques such as DL-GAN, DL-CNN, FR-CNN, CNNSVM, and SVM algorithms achieved low kappa values such as 0.984, 0.975, 0.982, 0.968, and 0.983 correspondingly. From the above discussed results, it is apparent that the proposed BDA-CSODL technique has accomplished effectual medical image classification performance.
In this study, a new BDA-CSODL technique is presented for biomedical image classification in Apache Spark environment. BDA-CSODL technique encompasses BF-based preprocessing, Otsu-based segmentation, Inception v3-based feature extraction, SGD-based hyperparameter optimization, LSTM-based classification, and CSO-based parameter tuning. Both SGD and CSO algorithms are designed in such a way that it considerably boosts medical image classification performance. In order to showcase the superior performance of BDA-CSODL approach, a comprehensive experimental analysis was carried out against benchmark medical images. The obtained experimental results showcase the supremacy of the proposed BDA-CSODL technique over other recent techniques under different performance measures. In future, multi-modal fusion-based medical image classification techniques can be designed to accomplish enhanced classification results.
The author extends his appreciation to the Deanship of Scientific Research at Majmaah University for funding this study under Project Number (R-2022-61).