3.1. Classifying blood cells with deep convolutional neural networks

An important part of the data acquired by blood tests is the number of white blood cells (WBCs) or leukocytes, usually differentiated into total and differential WBC count, where the latter describes the absolute and relative numbers of WBC subtypes (neutrophils, lymphocytes, basophils, eosinophils, and monocytes) in the sample. The amount of WBCs in the sample provides information on the state of the patient's innate and adaptive immune system, e.g., a significant changes in the WBC count relative to the patient's baseline is evidence that their body is being affected by an antigen, whereas variations in the specific WBC subtypes can correlate with specific types of antigens or different pathways of immune and inflammatory reaction. Therefore, detailed measurement and understanding of the WBC counts is an important part of the quantitative picture of health and the organism's general condition.

the high accuracy of 98.6%. While the presently used model performs less well (accuracy of 86%) when classifying WBCs into multiple individual type categories as opposed to binary classification, given the high performance and simplicity of this purely software-based approach, Athelas team plans to extend it to more complex problems, including datasets containing other cell types, which could enable faster improvement cycles, increased accessibility, and

A Review on Machine Learning and Deep Learning Techniques Applied to Liquid Biopsy

http://dx.doi.org/10.5772/intechopen.79404

53

During the last decade, human ageing research has received an increasing amount of mainstream interdisciplinary attention [19, 20], with an emerging tendency to approach various

Insilico team developed a DL system designed to predict human chronological age from biochemical data obtained from a basic blood test [21], narrowing an extensive set of potential ageing-related biomarkers to a limited subset of the most salient ones. A dataset of <sup>&</sup>gt; <sup>6</sup> � 104 records was used, with each record consisting of a patient's age, sex, and 46 blood biochemical markers. The dataset was preprocessed, normalising all blood marker values to 0–1 range, and

An ensemble of 21 feed-forward deep neural networks (DNNs) was created as the ML model, with a range of values assigned to DNN parameters such as the number of hidden layers, the number of processing units per layer, activation function, and optimization and regularisation methods. The permutation feature importance method [22] was used to evaluate the relative importance of the various biochemical markers with regard to ensemble accuracy. Batch normalisation [23] was used to reduce the effects of overfitting and increase the stability of

The best results were obtained from a DNN with five hidden layers, using regularised mean squared error (MSE) function as the loss function, parametric rectified linear unit (PReLU) [24] activation function in each layer, and AdaGrad [25] optimiser of the loss function. The highestscoring DNN performed with 82% ε-prediction accuracy at ε ¼ 10 (i.e., considering the sample as correctly recognised if the predicted age is �10 years of the true age), out-performing several classes of competing ML models. Multiple models for combining individual DNNs into an ensemble (stacking) were evaluated, with the best being the elastic net model [26]. The most important blood markers were discovered to be albumin, glucose, alkaline phosphatase,

Insilico team created an online service (http://www.aging.ai) to make the DNN ensemble available to the general public, allowing patients to use their blood test data to evaluate the age prediction system and serving as a proof of concept for estimating ageing-related variables using readily available biochemical data. Additional data sources, including transcriptomic and metabolomic markers from liquid and individual organ biopsies, as well as imaging data, are being considered. Insilico team suggests that similar systems could also be developed for model organisms in order to perform cross-species analysis of individual biological markers

and their importance in predicting both chronological and biological age.

better patient outcomes, compared to previously used methods of cell count analysis.

3.2. Using deep neural networks for detection of ageing-related biomarkers

aspects of the natural ageing process as potentially treatable conditions.

then split into training and test datasets with ratio 90:10.

convergence of the models.

urea, and erythrocyte count.

Traditional methods of estimating the WBC count generally fall into one of two categories manual and automated. The historical manual inspection of the blood sample involved counting the number of cells in a blood sample under a microscope and extrapolating under the assumption of uniform cell distribution across the entire bloodstream. Automated methods involve specialised equipment such as Coulter counters [14] or laser flow cytometers [15] which can provide accurate results and good performance [16] but are generally expensive and require specialised training to operate.

In this light, the ML-based approach provides a potential improvement over the aforementioned techniques due to several reasons. First, it requires far less expensive equipment due to being built around simple imaging solutions. Furthermore, unlike earlier methods, it is able to provide almost instantaneous results after the initial training stage. Finally, its performance can be expected to improve over time, in proportion to growing dataset sizes and, being mostly software-based, it can be expanded and advanced continually and "over the air", without requiring extensive changes in the underlying infrastructure.

We illustrate this approach using an example problem provided by Athelas team [17], namely, binary classification of a stained image of a WBC as either polymorphonuclear or mononuclear.1 The training dataset consisted of hand-labelled images of stained WBCs of all given types in various proportions. Before the dataset could be used, several preprocessing steps were taken, including removing images with multiple cells and using transformations such as flips and rotations in order to increase the size and variability of the training dataset. By using transformed versions of the images, the training dataset size was increased from approximately 350 to 104 .

For the ML model, Athelas team used the LeNet-5 [18] convolutional neural network (CNN) [3, 4, 9] due to its simplicity and availability. The model was tested against a test dataset of 71 images (20% of the original training set and 0.7% of the training set after transformations), with

<sup>1</sup> Eosinophils, basophils, and neutrophils are polymorphonuclear, while lymphocytes and monocytes are mononuclear.

the high accuracy of 98.6%. While the presently used model performs less well (accuracy of 86%) when classifying WBCs into multiple individual type categories as opposed to binary classification, given the high performance and simplicity of this purely software-based approach, Athelas team plans to extend it to more complex problems, including datasets containing other cell types, which could enable faster improvement cycles, increased accessibility, and better patient outcomes, compared to previously used methods of cell count analysis.

### 3.2. Using deep neural networks for detection of ageing-related biomarkers

3. Using machine learning techniques in blood tests

and require specialised training to operate.

mately 350 to 104

52 Liquid Biopsy

1

.

requiring extensive changes in the underlying infrastructure.

3.1. Classifying blood cells with deep convolutional neural networks

An important part of the data acquired by blood tests is the number of white blood cells (WBCs) or leukocytes, usually differentiated into total and differential WBC count, where the latter describes the absolute and relative numbers of WBC subtypes (neutrophils, lymphocytes, basophils, eosinophils, and monocytes) in the sample. The amount of WBCs in the sample provides information on the state of the patient's innate and adaptive immune system, e.g., a significant changes in the WBC count relative to the patient's baseline is evidence that their body is being affected by an antigen, whereas variations in the specific WBC subtypes can correlate with specific types of antigens or different pathways of immune and inflammatory reaction. Therefore, detailed measurement and understanding of the WBC counts is an impor-

tant part of the quantitative picture of health and the organism's general condition.

Traditional methods of estimating the WBC count generally fall into one of two categories manual and automated. The historical manual inspection of the blood sample involved counting the number of cells in a blood sample under a microscope and extrapolating under the assumption of uniform cell distribution across the entire bloodstream. Automated methods involve specialised equipment such as Coulter counters [14] or laser flow cytometers [15] which can provide accurate results and good performance [16] but are generally expensive

In this light, the ML-based approach provides a potential improvement over the aforementioned techniques due to several reasons. First, it requires far less expensive equipment due to being built around simple imaging solutions. Furthermore, unlike earlier methods, it is able to provide almost instantaneous results after the initial training stage. Finally, its performance can be expected to improve over time, in proportion to growing dataset sizes and, being mostly software-based, it can be expanded and advanced continually and "over the air", without

We illustrate this approach using an example problem provided by Athelas team [17], namely, binary classification of a stained image of a WBC as either polymorphonuclear or mononuclear.1 The training dataset consisted of hand-labelled images of stained WBCs of all given types in various proportions. Before the dataset could be used, several preprocessing steps were taken, including removing images with multiple cells and using transformations such as flips and rotations in order to increase the size and variability of the training dataset. By using transformed versions of the images, the training dataset size was increased from approxi-

For the ML model, Athelas team used the LeNet-5 [18] convolutional neural network (CNN) [3, 4, 9] due to its simplicity and availability. The model was tested against a test dataset of 71 images (20% of the original training set and 0.7% of the training set after transformations), with

Eosinophils, basophils, and neutrophils are polymorphonuclear, while lymphocytes and monocytes are mononuclear.

During the last decade, human ageing research has received an increasing amount of mainstream interdisciplinary attention [19, 20], with an emerging tendency to approach various aspects of the natural ageing process as potentially treatable conditions.

Insilico team developed a DL system designed to predict human chronological age from biochemical data obtained from a basic blood test [21], narrowing an extensive set of potential ageing-related biomarkers to a limited subset of the most salient ones. A dataset of <sup>&</sup>gt; <sup>6</sup> � 104 records was used, with each record consisting of a patient's age, sex, and 46 blood biochemical markers. The dataset was preprocessed, normalising all blood marker values to 0–1 range, and then split into training and test datasets with ratio 90:10.

An ensemble of 21 feed-forward deep neural networks (DNNs) was created as the ML model, with a range of values assigned to DNN parameters such as the number of hidden layers, the number of processing units per layer, activation function, and optimization and regularisation methods. The permutation feature importance method [22] was used to evaluate the relative importance of the various biochemical markers with regard to ensemble accuracy. Batch normalisation [23] was used to reduce the effects of overfitting and increase the stability of convergence of the models.

The best results were obtained from a DNN with five hidden layers, using regularised mean squared error (MSE) function as the loss function, parametric rectified linear unit (PReLU) [24] activation function in each layer, and AdaGrad [25] optimiser of the loss function. The highestscoring DNN performed with 82% ε-prediction accuracy at ε ¼ 10 (i.e., considering the sample as correctly recognised if the predicted age is �10 years of the true age), out-performing several classes of competing ML models. Multiple models for combining individual DNNs into an ensemble (stacking) were evaluated, with the best being the elastic net model [26]. The most important blood markers were discovered to be albumin, glucose, alkaline phosphatase, urea, and erythrocyte count.

Insilico team created an online service (http://www.aging.ai) to make the DNN ensemble available to the general public, allowing patients to use their blood test data to evaluate the age prediction system and serving as a proof of concept for estimating ageing-related variables using readily available biochemical data. Additional data sources, including transcriptomic and metabolomic markers from liquid and individual organ biopsies, as well as imaging data, are being considered. Insilico team suggests that similar systems could also be developed for model organisms in order to perform cross-species analysis of individual biological markers and their importance in predicting both chronological and biological age.

## 3.3. Machine learning-based approach to Alzheimer disease biomarker discovery

In their study, Smalheiser team has developed [27] a ML-based model for predicting Alzheimer disease (AD) status of individual samples with high accuracy, using miRNAs and other small RNAs extracted from circulating exosomes obtained from liquid biopsy (blood plasma) samples.

The CNN received normalised 40 � 40 pixel images as input. They were passed to a layer of 6 convolutional filters with the size of 5 � 5, followed by a max-pooling layer in order to extract

where ð Þ p; q —pixel coordinates, y—input map, z—output map. This layer was followed by another convolutional filter layer, consisting of 12 filters, and, subsequently, by another maxpooling layer. The last layer was fully connected to the output layer by way of dot product between the weight and input vectors, passed to the sigmoid function which maps the values to the ½ � �1; 1 range. The filter parameters, network bias terms, and weight matrices were

Mao team compared their CNN-based classifier to a simpler, SVM-based method that depended on hand-crafted feature sets. Using the F-score (harmonic mean of precision and recall scores) as the comparison metric, they found that, after two rounds of five iterations, the F-score of the CNN-based classifier was 0.97, by 18.6 points exceeding the F-score (0.784) of the SVM-based classifier and hand-crafted feature set. They concluded that the CNN-based classifier presents a promising development towards automated CTC detection in images taken from blood samples, and that the technique could be adapted for use with microfluidics-

2�pþm,2�qþn

A Review on Machine Learning and Deep Learning Techniques Applied to Liquid Biopsy

n o, (1)

http://dx.doi.org/10.5772/intechopen.79404

55

the local signal in every 2 � 2 pixel region, defined by the max-pooling function,

<sup>0</sup> <sup>≤</sup> m, <sup>n</sup> <sup>≤</sup> <sup>2</sup> yi

p, <sup>q</sup> ¼ max

automatically adjusted by backpropagation with learning rate set to 0:1.

based liquid biopsy platforms for early diagnosis and monitoring.

methods

dient (SCG)) [38].

2

respectively.

4. Cancer detection and monitoring using neural network-based

4.1. Using artificial neural networks for lung cancer detection and diagnosis

Goryński team describes [36] an artificial neural network (ANN)-based model class used for early detection and diagnosis of lung cancer. In their study, a dataset consisting of a wide range of biochemical parameters obtained from blood samples, as well as results from medical interviews (48 values in total) from 193 patients of mixed age and sex was used to train a family of 10 multilayer perceptron network (MLP) [3, 4] architectures, using a range of activation functions (linear, logistic, and tanh) for both hidden and output layers, as well as varying number of processing units ("neurons") in the hidden layer and different training algorithms (gradient descent, Broyden-Fletcher-Goldfarb-Shanno (BFGS) [37], and scaled conjugate gra-

Goryński team found that two of the trained models, named MLP 48–9-22 (trained using BFGS algorithm and using linear and tanh activation functions for hidden and output layers, respectively) and MLP 48–15-2 (SCG algorithm, logistic and tanh activation functions) gave highly

The naming scheme represents the number of "neurons" in the input, hidden, and output layers of the MLP model,

zi

A sample set of N ¼ 70 was used to construct the training dataset consisting of normalised miRNA expression data across 465 loci. Cross-validation was used in feature selection to evaluate the impact of values from specific loci as features. The samples were randomly divided into 7 partitions of 5 positive and 5 negative samples each and cross-validation was performed on these partitions, using 6 partitions for training and 1 for evaluation. The random partitioning was repeated 10 times in order to acquire 70 estimate points of the performance measures of interest, one for each sample in the set. These values were averaged and their relative performance was assessed using area under the curve of the receiver operating curve (ROC), Matthews correlation coefficient (MCC) [28], and F1 score.

Smalheiser team evaluated three different ML classifier algorithms—C4.5 decision trees [29] (using the J48 implementation), support vector machines (SVMs) [30], and adaptive boosting (AdaBoost) [31]. After selecting 50 most significant features, as per Mann-Whitney U test [32], the C4.5 classifier produced the best results, based on which it was selected as the feature selection method. The feature significance was measured by the number of times the given miRNA locus was used as a node in the decision tree over the 70 runs. The 18 highest-scoring features were selected to move on to the next step. AdaBoost algorithm was used for the final feature selection from the set of 18 features, producing an optimised set of 7 features which were then used with all 70 data samples to produce the final dataset.

The best model used by Smalheiser team was able to correctly classify, on average, 29 out of 35 samples from the AD group and 31 out of 35 samples from the control group, yielding accuracy in the range of 83–89%. Smalheiser team concluded that ML-based classifiers are able to produce highly accurate predictions of AD occurrence, using a dataset of only 7 miRNAs and that integrating exosome miRNA data with other data is likely to further increase performance of these models.

## 3.4. Detection and classification of circulating tumour cells using machine learning methods

The presence of circulating tumour cells (CTCs) in blood samples indicates the tumour response to chemotherapeutic drugs and contributes to the mechanism for subsequent growth of derived tumours (metastasisation) in distant tissues. Evaluation of CTCs can yield the diagnosis or help to follow the tumour response to chemotherapeutic drugs.

Mao team designed a deep (six layers) CNN for image-based circulating tumour cell detection with automatically learned network parameters [33]. They used a dataset of 45 phase contrast microscopy [34, 35] images, of which 35 randomly selected images were used for training and the remaining 10 for testing the network. The experiment was repeated 5 times in order to minimise network bias.

The CNN received normalised 40 � 40 pixel images as input. They were passed to a layer of 6 convolutional filters with the size of 5 � 5, followed by a max-pooling layer in order to extract the local signal in every 2 � 2 pixel region, defined by the max-pooling function,

3.3. Machine learning-based approach to Alzheimer disease biomarker discovery

(ROC), Matthews correlation coefficient (MCC) [28], and F1 score.

were then used with all 70 data samples to produce the final dataset.

samples.

54 Liquid Biopsy

mance of these models.

minimise network bias.

methods

In their study, Smalheiser team has developed [27] a ML-based model for predicting Alzheimer disease (AD) status of individual samples with high accuracy, using miRNAs and other small RNAs extracted from circulating exosomes obtained from liquid biopsy (blood plasma)

A sample set of N ¼ 70 was used to construct the training dataset consisting of normalised miRNA expression data across 465 loci. Cross-validation was used in feature selection to evaluate the impact of values from specific loci as features. The samples were randomly divided into 7 partitions of 5 positive and 5 negative samples each and cross-validation was performed on these partitions, using 6 partitions for training and 1 for evaluation. The random partitioning was repeated 10 times in order to acquire 70 estimate points of the performance measures of interest, one for each sample in the set. These values were averaged and their relative performance was assessed using area under the curve of the receiver operating curve

Smalheiser team evaluated three different ML classifier algorithms—C4.5 decision trees [29] (using the J48 implementation), support vector machines (SVMs) [30], and adaptive boosting (AdaBoost) [31]. After selecting 50 most significant features, as per Mann-Whitney U test [32], the C4.5 classifier produced the best results, based on which it was selected as the feature selection method. The feature significance was measured by the number of times the given miRNA locus was used as a node in the decision tree over the 70 runs. The 18 highest-scoring features were selected to move on to the next step. AdaBoost algorithm was used for the final feature selection from the set of 18 features, producing an optimised set of 7 features which

The best model used by Smalheiser team was able to correctly classify, on average, 29 out of 35 samples from the AD group and 31 out of 35 samples from the control group, yielding accuracy in the range of 83–89%. Smalheiser team concluded that ML-based classifiers are able to produce highly accurate predictions of AD occurrence, using a dataset of only 7 miRNAs and that integrating exosome miRNA data with other data is likely to further increase perfor-

3.4. Detection and classification of circulating tumour cells using machine learning

diagnosis or help to follow the tumour response to chemotherapeutic drugs.

The presence of circulating tumour cells (CTCs) in blood samples indicates the tumour response to chemotherapeutic drugs and contributes to the mechanism for subsequent growth of derived tumours (metastasisation) in distant tissues. Evaluation of CTCs can yield the

Mao team designed a deep (six layers) CNN for image-based circulating tumour cell detection with automatically learned network parameters [33]. They used a dataset of 45 phase contrast microscopy [34, 35] images, of which 35 randomly selected images were used for training and the remaining 10 for testing the network. The experiment was repeated 5 times in order to

$$\mathbf{z}\_{p,q}^i = \max\_{0 \le m\_\nu \le 2} \left\{ y\_{2 \times p + m, 2 \times q + n}^i \right\} \tag{1}$$

where ð Þ p; q —pixel coordinates, y—input map, z—output map. This layer was followed by another convolutional filter layer, consisting of 12 filters, and, subsequently, by another maxpooling layer. The last layer was fully connected to the output layer by way of dot product between the weight and input vectors, passed to the sigmoid function which maps the values to the ½ � �1; 1 range. The filter parameters, network bias terms, and weight matrices were automatically adjusted by backpropagation with learning rate set to 0:1.

Mao team compared their CNN-based classifier to a simpler, SVM-based method that depended on hand-crafted feature sets. Using the F-score (harmonic mean of precision and recall scores) as the comparison metric, they found that, after two rounds of five iterations, the F-score of the CNN-based classifier was 0.97, by 18.6 points exceeding the F-score (0.784) of the SVM-based classifier and hand-crafted feature set. They concluded that the CNN-based classifier presents a promising development towards automated CTC detection in images taken from blood samples, and that the technique could be adapted for use with microfluidicsbased liquid biopsy platforms for early diagnosis and monitoring.
