Diagnosis of dengue virus infection using spectroscopic images and deep learning

Study design and sample size selection

One of the important steps in the process of designing an empirical study is to select an appropriate sample size for clinical data. If sample size is insufficient or not selected carefully, the study might not be discovered underlying some important treatment effects. An optimal sample size selection saves personnel effort, time, and cost. Sample size must be selected and justified based on statistical calculations. A statistically selected sample size plays crucial role for validation and trustworthiness of the experiments.

For this study design, number of samples are determined using power calculations in G*Power 3.1.9.6 (Faul et al., 2007). Statistical power depends on different factors such as effect size, sample size, and the decision criterion (α-value). We used the criteria of 0.20, 0.50, and 0.80 provided by Cohen for small, medium, and large effect sizes, respectively (Cohen, 1992). For this purpose, statistical t-test with difference between two independent means are employed. The α error probability (false positive result or type-I error) is set equal to 0.05 and Power (1−β error prob.) that a false negative result or type-II error is considered 0.95. Figure 2 depicts the detectable total sample size as a function of required power (1−β error prob.). Statistical power calculations revealed that this study can be design using minimum sample size of 85 with actual power of 0.952, and minimal effect size of 0.8. However, using this criterion, we have chosen 100 samples (60 Dengue positive and 40 negative samples). The current sample size is sufficient to design this study for DENV infection detection to capture underlying chemical signatures for treatment and rehabilitation of the patients.

Human blood sera and sample preparation

On the basis of the experimental study design (‘Method’), 60 DENV infected and 40 non-infected blood samples were obtained from different hospitals of Rawalpindi, Pakistan during dengue epidemic outbreak in 2019. Verbel consent of volunteer patients was obtained and approved by the Research and Ethical Committee of Rawalpindi Medical College (RMC) and Allied Hospitals (Holy Family Hospital, Benazir Bhutto Hospital, District Headquarters Hospital, Rawalpindi), Pakistan. Collected samples were kept in freezer and thawed before recording Raman spectra. For acquisition of Raman spectra, there is no need of any special preparation.

Raman spectra collection

The spectra of sera samples were collected using PeakSeeker Pro-785 Raman system with RSM-785 microscope. The associated diode laser @785 nm was used for acquiring of Raman spectra at resolution of 10 cm⁻¹. To focus the laser on sera sample, a 10X objective lens was employed with back-scattering configuration to obtain Raman spectra at the 20 µl quantity placed at aluminum holder. The collected Raman spectra lies in the range of 200–2,000 cm⁻¹. Owing to the range 540–1,830 cm⁻¹ which contains notable spectral variations is chosen for analysis of spectra.

The presence of additive noise makes it difficult to analyze the recorded spectra at band of interest with naked eyes. Spectral fluorescent of closely bands is visualized and analyzed for annotation. Prior to analysis, it is imperative to reduce the effect of noisy bands in the Raman spectra. All these spectra were pre-processed using Matlab2014a for noise reduction, reference point normalization and correction to make the bands of interest and groups recognized precisely. The polynomial of degree 5 with 13-point window size is used for smoothing by employing Savitzky–Golay method. The fluorescence backgrounds of all spectra were corrected by the removal of baseline.

Dataset of Raman images

In this study, we have a dataset of 100 subjects (60 infected and 40 non-infected healthy subjects). For each subject, 20 Raman spectra images were acquired to prevent any experimental artifacts, human error, and for an unbiased statistical analysis. These spectra were recorded from different positions by shining a laser power of 150 mW for 10 s. Such 20 different spectra of a subject are reasonably different from others. These images are varying in intensities while keeping the same molecular changes. In terms of intensity variation, all 2,000 Raman spectra are different from each other. Out of 2,000 spectra 1,200 are infected and 800 are healthy (Naseer et al., 2019b; Saleem et al., 2020).

Pre-processing

The only single pre-processing step is to resize the Raman spectra for compatibility of ResNet101 input of size 224  × 224 by employing TL concept. The whole dataset was resized and fed it to the modified ResNet101 for training and testing. The dataset is randomly divided into 67%:33% ratio for model development and testing respectively, as shown in Fig. 3. In first step, modified ResNet101 deep learning model is developed on 67% of dataset. After successful training, the TL-ResNet101 model is evaluated on independent 33% unseen dataset which is not used during model training. All experiments are performed on Intel Xeon E-2246G, 3.6 GHz processor using NVIDIA GeForce RTX-2080 GPU.

Convolutional neural networks (CNN)

Owing to precision and scalability, CNN gains popularity among researchers. It is being widely used in several real-life applications such as, computer vision, pattern recognition, medical disease diagnosis and natural language processing. CNN has ability to solve complex and non-linear classification problems by extracting most discriminant image features using cascading layers (LeCun et al., 1998). Several convolutional, and pooling layers are stacked to extract most influential features followed by the classification layers.

TL is brilliant idea to use already learned weights of deep neural networks and modify some of the network layers for new classification problems. In such scenario, the modified fully connected layers of the networks are re-train on new dataset. The ResNet101 deep learning model is trained on ImageNet dataset for 1,000 categories. In this research, the modified ResNet101 network is trained for binary class instead of 1,000 classes for DENV infection diagnosis using Raman spectroscopic images.

To solve the DENV infection diagnosis, ResNet101 (He et al., 2016) is modified by employing TL concept. DENV infection diagnosis using human blood Raman spectra has not yet explored using Deep TL approach. The thematic diagram of modified ResNet101 by employing TL concept is shown in Fig. 4. It is highlighted in Fig. 4 that last three layers of ResNet101 are modified and trained for DENV infection diagnosis. TL and its associated CNN concepts are explained in the following sub-sections.

Pooling layers

These layers are used to reduce image size by down sampling for translation invariant and decreasing computational cost through dimension reduction. For object identification, max pooling is a well-known operation to be performed (Yamashita et al., 2018; Zhao et al., 2018). The result of max pool y_i is obtained by the following equation: (2)${\displaystyle y_i=\underset{1\le j\le M\times M}{max}\left(x_j\right),x_j\in {\mathbf{X}}_i}$

where $\mathbf{X}=\left\{{\mathbf{X}}_0,{\mathbf{X}}_1,..,{\mathbf{X}}_n\right\}$ combines local areas X₀, X₁, .., X_n of input image. For instance, for i ^th sub-image ${\mathbf{X}}_i=\left(x_1,x_2,..,x_{M\times M}\right)$ each element of x have Msize. The pooling operation identifies the discriminant image features used DENV identification from Raman spectra. We used a kernel size 2  ×  2 of stride S = 2.

Cost function

It assesses the difference between target and networks predicted output in feed-forward fashion. Cross entropy is utilized as cost function for deep neural networks training. Generally, multi class cost function is defined in Eq. (3). (3)${\displaystyle LF=-\frac{1}{N}\sum _{i=1}^N{z}_i.log\left(p\left(z_i\right)\right)+\left(1-z_i\right).log\left(1-p\left(z_i\right)\right)}$ For our binary classification problem, Eq. (3) is modified as follows: (4)${\displaystyle L{F}_{Binary}=-z_i.log\left(p\left(z_i\right)\right)+\left(1-z_i\right).log\left(1-p\left(z_i\right)\right)}$

where, z and $p\left(z\right)$are output class labels and corresponding probability.

FC layers

The image features obtained from the ‘avg_pool’ layer of ResNet101 are fed to the FC layers to yield output. FC layers produce likelihood for each class generated by the output neurons (classes) which may vary for problem to problem. Finally, softmax function is employed to output neurons (two for our problem) and get predicted class of DENV-infected or non-infected subjects. The softmax function is defined below: (5)${\displaystyle soft\text{\_}max\left(\hat{Y}i\right)=\frac{exp\left(yi\right)}{\sum _{j=1}^n\left(yj\right)}.}$

Transfer learning (TL)

Design and development of a deep learning model from scratch required significantly large amount of annotated data. For relatively small dataset, Transfer Learning (TL) is an alternative and popular idea to exploit the pre-learned weights of deep learning models. In this way, the partial knowledge of previously trained model is utilized to solve multiple new classification problems. Principally, TL based deep learning network is to start with pre-initialized weights from training of other related problem and accordingly required less training time. In TL, the existing architectures are modified, and new layers are added for the solution of new problem. In the proposed DENV infection diagnosis, we modified ResNet101 deep learning model by employing TL concept. Three layers of the existing ResNet101 are removed and new layers are added and trained for the diagnosis of DENV infection.

Recently, TL concept gains popularity among researchers that exploit the networks learned weights and used for entirely new classification problem. In TL, the networks weights are frozen except few of the last layers (Russakovsky et al., 2015; Xie et al., 2015). Generally, development of DNN models from scratch requires enormous amount of annotated data. Inherently, medical images labelled data is very small in amount for Deep learning model development. To cope this, it is suitable to modify existing models and apply TL concept on relatively small dataset. TL is employed successfully to classify colon polyps, thyroid, skin cancer, lung nodule, COVID etc. (Esteva et al., 2017; Shin et al., 2016; Nayak et al., 2021).

The original ResNet101 was trained for 1,000 classes and having last three layers of ResNet101 namely ‘FC_1000’, ‘FC1000_Softmax’ and ‘FC_Classification’. We modified ResNet101 by replacing the mentioned layers with new layers of ‘FC_2’, ‘FC2_Softmax’ and ‘Class_output’ for our binary classes ‘Infected’, and ‘Healthy’ as depicted in Fig. 4. The ‘avg_pool’ layer of the TL-ResNet101 having 2,048 neurons connected with succeeding layers of two neurons each representing a class. The replaced layers produced high level abstract pattern representation for effective DENV-infection identification.

In training process, hyper-parameters of the networks are tuned for improved performance. The TL-ResNet101 parameters are set empirically as: learning rate of 0.0001, batch size of 20, data augmentation of (−30, 30) and epochs limit 600 to avoid overfitting and performed generalization. Owing to the residual learning capability and better generalization, ResNet101 architecture is selected that have advantage over the other architectures such as AlexNet and VGG (Simonyan & Zisserman, 2014; Krizhevsky, Sutskever & Hinton, 2012). The architectural diagram of residual learning is shown in Fig. 5.

In Fig. 5, x and F(x) are the input and its corresponding activation function. When the activation function produces $F\left(x\right)=0$ the residual process returned an identity y = x. This process reduces the over-fitting risk hence get better generalization. ResNet101 architecture detail is given in He et al. (2016).

Performance assessment

Performance of the developed training and testing models are evaluated using quality measures of accuracy, MCC, F-score ROC (true positive rate vs false positive rates), sensitivity, specificity, Kappa index and AUC. These measures are also used to compare the developed approach performance over the other models on the similar dataset. Mathematical definitions of quality measures are given below:

(6)${\displaystyle Accuracy=\frac{TP+TN}{P+N}}$ (7)${\displaystyle Senstivity\left(orrecall\right)=\frac{TP}{P}}$ (8)${\displaystyle Specificity=\frac{TN}{N}}$ (9)${\displaystyle F\text{\_}Score=2\times \frac{precision\times senstivity}{precision+senstivity}}$ (10)${\displaystyle Precision=\frac{TP}{TP+FP}}$ (11)${\displaystyle MCC=\frac{TP\times TN-FP\times FN}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}}$

where, P, N FP, FN, and TP represent, total positive, total negative, False Positive, False Negative, and True Positive, respectively.

Experimental results

The performance of proposed DENV-TLDNN is assessed via extensive experiments for diagnosis of DENV infection using Raman Spectroscopic images. The dataset used in this research is obtained from the real patient’s blood samples. The developed model is evaluated by various quality measures and is compared with other state-of-the-art techniques. Results indicated that DENV-TLDNN outperformed over the previously developed approaches. The developed model has potential to use for DENV infection diagnosis at massive scale. Training and testing data randomly split into 67%:33% ratio for model development and evaluation as indicated in Fig. 3. Data splitting presumed that it follows sampling criteria that it marginally distinct the overall data.

Training and testing accuracies along with loss curves of the DENV-TLDNN are shown in Figs. 6 and 7, respectively. Both curves show that model learns parameters consistently and efficiently. Similarly, the average training and testing performance is shown in Fig. 8. Performance comparison in terms of ROC curves of DENV-TLDNN and TL-ResNet50 is shown in Fig. 9. Moreover, AUC values of the developed and TL-ResNet50 are 0.965 and 0.901 respectively. Table 1 highlights the various performance comparisons of developed approach with TL-ResNet50, and SVM. It is observed from Table 1 that the proposed approach outperformed others. The computational complexity comparison of the DENV-TLDNN and other approaches is shown in Table 2.