The evolution of bone marrow morphology is necessary in Acute Myeloid Leukemia (AML) prediction. It takes an enormous number of times to analyze with the standardization and inter-observer variability. Here, we proposed a novel AML detection model using a Deep Convolutional Neural Network (D-CNN). The proposed Faster R-CNN (Faster Region-Based CNN) models are trained with Morphological Dataset. The proposed Faster R-CNN model is trained using the augmented dataset. For overcoming the Imbalanced Data problem, data augmentation techniques are imposed. The Faster R-CNN performance was compared with existing transfer learning techniques. The results show that the Faster R-CNN performance was significant than other techniques. The number of images in each class is different. For example, the Neutrophil (segmented) class consists of 8,486 images, and Lymphocyte (atypical) class consists of eleven images. The dataset is used to train the CNN for single-cell morphology classification. The proposed work implies the high-class performance server called Nvidia Tesla V100 GPU (Graphics processing unit).

Anomalies bring about acute Myeloid Leukemia (AML) in the DNA (Deoxyribonucleic acid) that controls the improvement of cells in your bone marrow [

The proposed work is carrying out the terms as Adult AML is a type of cancer in bone marrow makes abnormal myeloblasts. Usually, it gets poorer when it is not treated. It is a type of acute leukemia in generally noted in adults [

Fast R-CNN uses a region proposal network (RPN) with the CNN. The RPN shares fully-connected convolutional highlights the detection network empowering about without region proposals. It is an entirely CN that simultaneously predicts object limits and adjacent Ness scores. R-CNN is prepared to finish creating the region proposals which utilizes for location. It is converged into solitary network by sharing convolutional highlights. The RPN part advises brought together network. Faster R-CNN comprises two modules: the primary module and the subsequent module. The primary module is completely a convolutional network that proposes regions and the subsequent module is the Fast R-CNN indicator that utilizes the anticipated regions [

In this work, the process with the high-class performance server called Nvidia DGX. It is a line of Nvidia produced servers and workstations specializing in GPGPU (General-purpose Graphics processing unit is a graphics processing unit to accelerate deep learning applications. The product line is provided to bridge the gap among AI and GPUs accelerators in device has certain features specializing it for DL workloads. DGX-1 servers feature 8 GPUs relies on Volta or Pascal daughter cards with HBM 2 memory, connected by an NV Link mesh network. The initial Pascal-based DGX-1 delivers 170 teraflops of precision processing while volta-based upgrade enhanced to 960 teraflops [

Recently, the rapid growth of artificial intelligence has carried new assumptions to the various fields like medicine, agriculture, industry automation, banking sector, risk management, and fraud detection, especially for diagnosing the diseases, finding the variance, prognostication, and visualization [

It is reported that approximately 20 000 new cases are predicted with AML, and a 12,000-mortality rate was noted in 2020. The development of AML is considered to be miserable, specifically for patients 65 years old, and the report says that 59% of new cases are found over a specific region every year and 73% AML-based deaths. With this population, the roughly one-year survival rate is

Some patients are diagnosed with AML and treated with IC followed by consolidation therapy and stem cell transplant. Lower intensity treatments like hypomethylating agents are available for some patients who may not been IC members for the past few decades [

NIC or determining candidacy (IC) is not so standardized. Some prevailing treatment guidelines are suggested against the age as a dominating factor. However, no consensus algorithm is proposed as in data [

The proposed work is carrying out the terms as Adult acute myeloid leukemia (AML) is a cancer type where bone marrow is identified with abnormal myeloblasts. There is no proper staging system for adult AML. The data is collected from the American Cancer Society’s for leukemia in US (2020). Roughly 60,530 leukemia cases and 23,100 deaths are predicted. Roughly 19,940 new cases of AML are identified. The sample is taken from the adults. Almost all will be adults. The following

This research considers BioGPS for AML which is characterized on molecular heterogeneity (

It is a technique that increases the number of trainings samples and eliminates over-fitting issues. Here, sample processing with augmentation is a step that is performed to enhance the dataset quality before it is adapted to a classifier model. It includes color corrections, orientation and resizing. Data augmentation is a manipulation applied to create various content versions to expose the model towards a more comprehensive array of training samples. The augmentation process is typically used for the training data. Therefore, transformation is measured as an augmentation that best-suited pre-processing step. In some cases, the provided images are composed of low contrast images. In the case of the prediction process, the low contrast images are not best suited. Therefore, the contrast adjustment is highly solicited. When the provided training data does not have a constant contrast level representation, it is less specific where constant contrast adjustment is highly appropriate. Indeed, random contrast altering during the training process gives better generalization. It is known as augmentation.

The proposed work is processed with the input layer as the various images

Layer (Type) | Output shape | Parameter |
---|---|---|

Input_1 (Input layer) | (None, 256, 256, 3) | 0 |

Block1_conv1 (Conv 2D) | (None, 256, 256, 64) | 1792 |

Block1_conv2 (Conv 2D) | (None, 256, 256, 64) | 36928 |

Block1_pool (Max Pooling 2D) | (None, 128, 128, 64) | 0 |

Block2_conv1 (Conv 2D) | (None, 128, 128, 128) | 73856 |

Block2_conv2 (Conv 2D) | (None, 128, 128, 128) | 147584 |

Block2_pool (Max Pooling 2D) | (None, 64, 64, 128) | 0 |

Block3_conv1 (Conv 2D) | (None, 64, 64, 256) | 295168 |

Block3_conv2 (Conv 2D) | (None, 64, 64, 256) | 590080 |

Block3_conv3 (Conv 2D) | (None, 64, 64, 256) | 590080 |

Block3_conv4 (Conv 2D) | (None, 64, 64, 256) | 590080 |

Block3_pool (Max Pooling 2D) | (None, 32, 32, 256) | 0 |

Block4_conv1 (Conv 2D) | (None, 32, 32, 512) | 1180160 |

Block4_conv2 (Conv 2D) | (None, 32, 32, 512) | 2359808 |

Block4_conv3 (Conv 2D) | (None, 32, 32, 512) | 2359808 |

Block4_conv4 (Conv 2D) | (None, 32, 32, 512) | 2359808 |

Block5_pool (Max Pooling 2D) | (None, 8, 8, 512) | 0 |

Flatten_1 (flatten) | (None, 1024) | 0 |

dense_1 (Dense) | (None, 1024) | 33555456 |

dropout_1 (Dropout) | (None, 1024) | 0 |

dense_2 (Dense) | (None, 1024) | 1049600 |

dropout_3 (Dense) | (None, 15) | 2050 |

The output shape is acquired as 1024 and zero parameters in flattening one. In dense one along with the zero parameters output 33555456. In dropout one, the parameters are zero to get the output size as 1024. In dense two along with the 1049600 parameters output attaint as 1024. In dropout, two parameters are taken as 2050 to get the output size as 15. The hyperparameter considers the 32 optimized batch size values in training the dataset. The dropout value is optimized as 0.2. The loss is analyzed with the categorical cross-entropy. The ReLu implies the activation function for convolution layers.

The proposed model provides a novel regional network with superior computational efficiency and performance. The proposed detection network is a single network that shares the convolutional features, and the convolutions are shared during the testing time. The anticipated model is proposed to predict the AML. The fully convolutional network indicates the objectless and object bounds at every position. The functionality of the network is initiated by considering the convolutional base network, which is fed as an input image, i.e., resized image. The convolutional layer’s output depends on the NN stride. The sliding window is utilized to generate the region proposal over convolutional feature mapping by the network layer (last). The proposed network is used during feature mapping to learn whether an object is provided and its location and dimensions in the input image. The anchors carry it out, and it is centered at the sliding window and related to the aspect ratio and scaling. The classifier layer is utilized to provide probabilities whether or not points over the feature map contains an object within the anchors at the point.

The input image is initially passed via the input network to attain feature mapping. The bounding box coordinates pool the features from the mapped feature. The ROI pooling layer performs it. This ROI layer helps to 1) measure the region related to the feature map; 2) partition the regions to the fixed output size. After passing it to the fully connected layers, the classification provides the features. The proposed model is trained independently, and the CNN is initialized with weights from the trained network and fine-tuned to perform region-based classification task. The anticipated network model is trained independently. The backbone CNN weight is initialized from the trained network for classification, and it is fine-tuned for the detection task. It does not share the convolutional layers. The network is initialized with weights from the Faster R-CNN model and fine-tuned for the regional task. In the final stage, the detector is fine-tuned, and the unique layers to the detector network are fine-tuned. The standard layers’ weights are fixed, and finally, convolutional layers share the same network to form a unified network model.

Researchers must assess the quality of the classification findings before evaluating the classifier’s performance. The classification performance is computed to graphically illustrate the viability of the newly designed algorithm, followed by a comparative study of existing classification algorithms and additional enhancements depending on their inadequacies. Some metrics like accuracy, sensitivity, specificity, and F1-score are commonly employed, and its execution principle will be momentarily introduced underneath.

It is also an error matrix to see if the classification result matches the original ground cover. It is the foundation for several additional assessment criteria. It is expressed in

The overall accuracy is calculated by dividing the number of accurately sampled data points by the entire sample. The formula for the computation is expressed in

The performance indicators are based on the confusion matrix, with TP denoting a lot of positive instances that are positive, TN representing the number of negative instances that are negative, FP indicating the series of adverse instances that are positive but are intended just to be positive, and FN indicating the number of positive instances that are negative but are designed to be positive. The model’s capacity to handle a real example positively or negatively is solely dependent on the accuracy of the land cover/land use forecast. The fraction of correctly predicted positive instances out of all optimistic predictions supplied by the predictor model is precision.

The recall is represented as the percentage of projected positive events that are always positive. It’s written like this in

The F1-score is a commonly used metric for classification issues. The recall and accuracy rates are calculated using the average harmonic technique, with the maximal value is 1 and the minimal value is 0.

Specificity is defined as the ability to correctly forecast samples that are not deemed valid, i.e., accurate negative samples.

The different metrics to evaluate the semantic segmentation model are Pixel accuracy, Intersection over union (IoU), Dice coefficient. The IoU is called as Jaccard index. The IoU is calculated as follows.

Dice coefficient: It is analyzed by the given below equation

The dice coefficient is similar to the IoU.

The following

Methods | Accuracy (%) | Precision (%) | Recall (%) | F-measure (%) | ROC | Error rate | MCC |
---|---|---|---|---|---|---|---|

Hybrid CNN | 95% | 80 | 77 | 90 | 86 | 0.089 | 0.4655 |

Standard CNN | 96% | 62 | 76 | 68 | 85.5 | 1.02 | 0.4535 |

CNN with boosting model | 93% | 84 | 77 | 73 | 84 | 1.56 | 0.4318 |

D-CNN with data augmentation | 94% | 83 | 76 | 78 | 80 | 1.63 | 0.4659 |

CNN with Genetic optimization | 94% | 92 | 96 | 95 | 78 | 0.077 | 0.4206 |

DCNN | 97 | 94 | 97 | 96 | 91 | 0.075 | 0.3565 |

98 | 0.080 | 0.2898 |

However, the DCNN model shows 94% precision. The recall of the anticipated faster R-CNN model is 98% which is 21%, 22%, 21%, 22%, 2% and 1% higher than other approaches. The F-measure of the anticipated model is 95.65% which is 5.65%, 27.65%, 22.65%, 17.65%, 0.65% higher than other approaches and 0.35% lesser than the DCNN model. The ROC of the anticipated ROC is 98% which is 12%, 11.5%, 14%, 18%, 20% and 7% superior to other methods. The error rate of the anticipated model is 0.080, and the other models are 0.089, 1.02, 1.56, 1.63, 0.077 and 0.075. Generally, the value of the MCC should be lesser than 1. When the model gives a lesser value than 1, then the model is said to be superior. The MCC of R-CNN is 0.2898, DCNN is 0.3565, CNN with GA is 0.4206, D-CNN with DA is 0.4659, CNN with boosting model is 0.4318, standard CNN is 0.4535, and hybrid CNN is 0.4655.

Dataset | ML algorithm | Accuracy (%) | Precision |
Recall |
F1-score | Time (min) | AUROC |
---|---|---|---|---|---|---|---|

Kaggle dataset for Leukemia classification | Hybrid CNN | 92.5 | 75 | 72 | 88 | 0.10 | 93.8 |

Standard CNN | 87.8 | 78 | 70 | 68 | 0.10 | 93.36 | |

CNN with boosting model | 88.9 | 72 | 72 | 71 | 0.78 | 94.86 | |

D-CNN with data augmentation | 86.7 | 75 | 73 | 78 | 6.80 | 92.40 | |

CNN with Genetic optimization | 87.8 | 86 | 90 | 93 | 1.100 | 93.32 | |

Standard CNN | 93.1 | 91 | 93 | 91 | 0.06 | 96.89 | |

Gene Expression dataset | Hybrid CNN | 90.1 | 74 | 70 | 89 | 0.05 | 92.50 |

Standard CNN | 91.1 | 76 | 69 | 69 | 0.08 | 92.14 | |

CNN with boosting model | 90.4 | 77 | 70 | 72 | 0.78 | 91.14 | |

D-CNN with data augmentation | 91.9 | 75 | 68 | 79 | 6.06 | 90.38 | |

CNN with Genetic optimization | 90.15 | 71 | 73 | 92 | 1.62 | 90.97 | |

Standard CNN | 93.25 | 72 | 93 | 90 | 0.04 | 93.56 | |

Bio GPS |
Hybrid CNN | 95 | 80 | 78 | 89.25 | 1.25 | 87 |

Standard CNN | 96 | 62 | 77 | 67 | 2.658 | 86.5 | |

CNN with boosting model | 93 | 84 | 78 | 72 | 4.56 | 85 | |

D-CNN with data augmentation | 94 | 82 | 77 | 77.9 | 3.56 | 82 | |

CNN with Genetic optimization | 94.50 | 92 | 96 | 94.24 | 4.57 | 80 | |

Standard CNN | 90.10 | 90 | 88 | 90 | 2.89 | 90 | |

While performing the analysis with the Gene expression dataset, the model provides 95.50%, 93%, 94%, 95.8%, and 94.58% for accuracy, recall, precision, F1-score, and AUROC. The accuracy of the anticipated model is 95.50% which is 5.4%, 4.4%, 5.1%, 3.6%, 5.35% and 2.25% higher than other approaches. The precision of the anticipated model is 93% which is 19%, 17%, 16%, 22%, and 21% higher than other approaches. The recall of the anticipated model is 94% which is 24%, 25%, 24%, 21%, 21% and 1% higher than other approaches. The F1-score of the anticipated model is 95.8% which is 6.8%, 26.8%, 23.8%, 16.8%, 3.8% and 5.8% higher than other approaches. The AUROC of the anticipated model is 94.58% which is 2.08%, 2.44%, 3.44%, 4.2%, 4.2%, 3.61% and 1.02% higher than other approaches. In the case of the Bio GPS dataset, the accuracy of the anticipated model is 97.30, which is 2.3%, 1.3%, 4.3%, 3.3%, 2.8% and 7.2% higher than other approaches.

The precision of the anticipated model is 94% which is 14%, 32%, 10%, 12%, 2% and 4% higher than other approaches. The recall of the anticipated model is 98% which is 20%, 21%, 20%, 21%, 2% and 10% higher than other approaches. The F1-score of the anticipated model is 95.65% which is 6.4%, 28.65%, 23.65%, 17.75%, 1.41% and 5.65% higher than other approaches. The AUROC of the anticipated model is 92% which is 5%, 4.5%, 7%, 10%, 12% and 2% higher than other approaches See

The research attempts to improve the diagnostic accuracy of AML. We proposed the novel AML detection model using a Deep CNN. The proposed Faster R-CNN models are trained with a Single-cell Morphological Dataset. Based on the experimental results, the pixel accuracy is 98.32% compared with the other CNN algorithms, producing better results.

Description | R-CNN | Fast R-CNN | U-Net | Mask R-CNN | Faster R-CNN |
---|---|---|---|---|---|

Pixel accuracy | 85.49 | 87.64 | 91.74 | 92.42 | 98.32 |

Intersection-over-Union (0–50) | 35.28 | 38.94 | 42.31 | 42.67 | 48.7 |

Dice coefficient | 83.74 | 86.23 | 90.85 | 92.17 | 97.56 |

Precise and accurate prediction of leukocytes is a complex task. The flexible structure of the nucleus generates a huge problem for the prediction of leukemia. A novel Faster R-CNN is proposed to overcome these issues. The optimal features are chosen to model a feature vector. Here, various metrics like hybrid CNN, standard CNN, boosting model, data augmentation, GA and DCNN model is evaluated with the proposed faster RCNN model. The proposed Faster R-CNN gives 97.30% prediction accuracy, 94% precision, 98% recall, 95.65% F-measure, 0.080 error rate and 0.2898 MCC. Other metrics like pixel accuracy, Intersection over Union and Dice coefficient are compared with other approaches like RCNN, Faster R-CNN, U-Net and mask R-CNN. The pixel accuracy is 98.32%, Intersection over union is 48.7%, and dice coefficient is 97.56%. The proposed model works well in the prediction of acute myeloid leukemia. However, the major limitation of the proposed model is the lack of optimized results. In the future, the proposed model is optimized with the meta-heuristic optimization approach to attain global and local optima.

I am extremely thankful to Management, Principal, Head of the department of Computer Science and Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology for providing me laboratory facility in Vel Tech-AI Research Centre to carry out this work in successful way.