A Data-Efficient Deep Learning Based Smartphone Application For Detection Of Pulmonary Diseases Using Chest X-rays
Hrithwik Shalu, Harikrishnan P, Akash Das, Megdut Mandal, Harshavardhan M Sali, Juned Kadiwala
AA Data-Efficient Deep Learning BasedSmartphone Application For Detection OfPulmonary Diseases Using Chest X-rays.
Hrithwik Shalu , Harikrishnan P , Akash Das ,Megdut Mandal , Harshavardhan M Sali , Juned Kadiwala August 21, 2020
1. University of Cambridge2. Indian Institute of Technology Madras3. Indian Institute of Technology Patna4. Kalinga Institute of Industrial Technology Bhubaneswar*corresponding author
Abstract
This paper introduces a paradigm of smartphone application baseddisease diagnostics that may completely revolutionize the way healthcareservices are being provided. Although primarily aimed to assist the prob-lems in rendering the healthcare services during the coronavirus pandemic,the model can also be extended to identify the exact disease that the pa-tient is caught with from a broad spectrum of pulmonary diseases. Theapp inputs Chest X-Ray images captured from the mobile camera whichis then relayed to the AI architecture in a cloud platform, and diagnosesthe disease with state of the art accuracy. Doctors with a smartphonecan leverage the application to save the considerable time that standardCOVID-19 tests take for preliminary diagnosis. The scarcity of trainingdata and class imbalance issues were effectively tackled in our approach bythe use of Data Augmentation Generative Adversarial Network (DAGAN)and model architecture based as a Convolutional Siamese Network withattention mechanism. The backend model was tested for robustness us-ing publicly available datasets under two different classification scenarios(Binary/Multiclass) with minimal and noisy data. The model achievedpinnacle testing accuracy of 99.30% and 98.40% on the two respectivescenarios, making it completely reliable for its users. On top of that asemi-live training scenario was introduced, which helps improve the appperformance over time as data accumulates. Overall, the problems of gen-eralizability of complex models and data inefficiency is tackled through themodel architecture. The app based setting with semi live training helpsin ease of access to reliable healthcare in the society, as well as help ineffective research of rare diseases in a minimal data setting. a r X i v : . [ ee ss . I V ] A ug ntroduction The increasing adoption of electronic technologies is widely recognized as acritical strategy for making health care more cost-effective. Smartphone-basedm-health applications have the potential to change many of the modern-daytechniques of how healthcare services are delivered by enabling remote diagno-sis [1], but it is yet to realize its fullest potential. There has been a paradigmshift in the research on medical sciences, and technologies like point-of-carediagnosis and analysis have developed with more custom-designed smartphoneapplications coming into prominence. Due to the high rate of infection, with thetotal number of confirmed cases exceeding twenty million since its recent out-break, COVID-19 was chosen as the initial disease target for us to study. Withstudies confirming that chest X-rays are irreplaceable in a preliminary screeningof COVID-19, we started with chest X-rays as the tool to detect the presence ofcoronavirus (COVID-19) in the patients.[2] Chest X-ray is the primary imagingtechnique that plays a pivotal role in disease diagnosis using medical imaging forany pulmonary disease. Classic machine learning models have been previouslyused for the auto-classification of digital chest images [3][4]. Reclaiming theadvances of those fields to the benefit of clinical decision making and computer-aided systems using deep learning is becoming increasingly nontrivial as newdata emerge[5][6][7], with Convolutional Neural Networks (CNNs) spearhead-ing the medical imaging domain [8]. A key factor for the success of CNNs is itsability to learn distinct features automatically from domain-specific images, andthe concept has been reinforced by transfer learning [9]. However, the process oflearning distinct features by standard supervised learning using ConvolutionalNeural Networks can be computationally non-efficient and data expensive. Theabove methods become incapacitated when combined with a shortage of data.Our approach represents a substantial conceptual advance over all other pub-lished methods by overcoming the problem of data scarcity using a one-shotlearning approach with the implementation of a Siamese Neural Network. Con-trasting to its counterparts, our method has the added advantage of being moregeneralizable and handles extreme class imbalance with ease. We leverage openchest X-Ray datasets of COVID-19 and various other diseases that were publiclyavailable (refer datasets section)[10]. Once a Siamese network has been tuned,it can capitalize on powerful discriminative features to generalize the predic-tive power of the network not just to new data, but to entirely new classesfrom unknown distributions[11][12]. Using a convolutional architecture, we canachieve reliable results that exceed those of other deep learning models withnear state-of-the-art performance on one-shot classification tasks. The world isbeing crippled by COVID-19, an acute resolved disease whose onset might resultin death due to massive alveolar damage and progressive respiratory failure [13].A robust and accurate automatic diagnosis of COVID-19 is vital for countries toprompt timely referral of the patient to quarantine, rapid intubation of severecases in specialized hospitals, and ultimately curb the spread. The definitivetest for SARS-CoV-2 is the real-time reverse transcriptase-polymerase chain re-action (RT-PCR) test. However, with sensitivity reported as low as 60-70%[14]nd as high as 95-97%[15], a meta-analysis concluded the pooled sensitivity ofRT-PCR to be 89%[16]. These numbers point out false negatives to be a realclinical problem, and several negative tests might be required in a single case tobe confident about excluding the disease[17]. A resource-constrained environ-ment demands imaging for medical triage to be restricted to suspected COVID-19 patients who present moderate-severe clinical features and a high pretestprobability of disease, and medical imaging done in an early phase might befeature deficient [18][19]. Although the cause of COVID-19 was quickly identi-fied to be the SARS-CoV-2 virus, scientists are still working around the clockto fully understand the biology of the mutating virus and how it infects hu-man cells[20]. All these calls for a robust pre-diagnosis method, which hopesto provide higher generalization, work efficiently with insufficient feature data,and tackles the problem of data scarcity. This is where our proposed methodof Data Augmentation Generative Adversarial Network (DAGAN) exploited bya Convolutional Siamese Neural Network with attention mechanism comes intothe picture, exhibiting a state of the art accuracy and sensitivity.
Generative Adversarial Networks (GANs) are deep learning based generativemodels which take root from a game theoretic scenario where two networks com-pete against each other as an adversary. The constituent network models – aGenerative Network and a Discriminative Network play a zero-sum game. GANarchitecture paved way for sophisticated domain-specific data augmentation bytreating an unsupervised problem as a supervised one, thus automatically train-ing the generative model.Figure 1: Figurative representation of data flow in a Generative AdversarialNetworkThe Generative Network utilizes the latent space, which is a projection orompression of the data distribution to generate plausible training examples.Latent space is the end result of mapping the points in the multidimensionalvector space to points in the problem domain. Random vectors drawn from datadistributions like the Gaussian distribution are used to seed the generative pro-cess. The Discriminative Network has the primary objective of classifying real(training set samples) and fake instances (generator samples). The Discrimina-tive network component of the GANs are usually normal binary classificationmodels, with real instances as positive examples and fake instances as negativeexamples. It connects to two loss functions, the generator loss and the dis-criminator loss. Training of discriminative networks penalizes the discriminatorloss with the examples from generative networks with constant weights and realexamples as its input, ignoring the generator loss. Similarly the generator istrained to create competent plausible samples and modifies its weights based onthe generator loss.
Data Augmentation procedure is crucial in the training procedure of a deeplearning model as it has proven to be an effective solution in tackling the prob-lem of overfitting at numerous occasions. As the data could be made moregeneralized, by providing the same with suitable augmentation strategies. Inthe case of images, Augmentation plays a crucial role. As to correctly identifyand recognise specific features in the same, a diverse set of considerably differ-ent sets of images are required. Image augmentation techniques are henceforthfound in diversely different ways, ranging from simple transforms (rotation) toadversarial data multiplicative methods such the one we would be using for ourpurposes, called the data augmentation generative adversarial networks (DA-GAN). Figure 2: The DAGAN architecturehe purpose and uniqueness of DAGAN when compared to other types ofGANs , is the ability to generate distinctive augmented images for any givenimage sample while preserving the distinctive class features intact. A generalnetwork architecture of the same is provided in Figure 2.
Generator : The Generator component of the DAGAN contains an encoderwhich provides a unique latent space representation for a given image and a de-coder which generates an image given a latent representation. Any given imageis first passed through the encoder to attain the corresponding latent represen-tation, to which a scaled noise ( usually sampled from a Gaussian distribution) is added to obtain a modified latent vector. The same is then passed throughthe decoder to obtain the corresponding augmented image.
Discriminator : The discriminator component of the DAGAN is similarto other GANs, where the basic purpose of which is to perform a binary clas-sification to tell apart the generated and real images. The discriminator takesas input a fake distribution (generated images) and a real distribution (Imagesbelonging to the same class).
Forming an ideal dataset for a typical multi-class classification task using stan-dard supervised learning methods is quite difficult. In addition to class imbal-ance issues, data for certain tasks such as medical image analysis could rarely becollected to meet ideal standards. One-shot learning methods helps tackle theseissues effectively. In the Deep Learning literature, Siamese Neural Networks aretypically used to perform one-shot learning.The Siamese Neural Network is a pair of neural networks trying to learndiscriminative features from a pair of data points from two different classes.Inour case the Siamese Networks would consist of two twin Convolutional Neu-ral Networks which accept distinct inputs but are joined together by an energyfunction.The latent vector is the overall output from either of the twin neu-ral networks, it is a unique and meaningful representation of individual imagespassed. In one shot learning the overall training objective is to obtain a vectorvalued function (Neural Network) which provides meaningful latent representa-tion vectors to the each image passed. As any machine learning task, the oneshot learning too has a loss function whose value conveys how close the networkis in attaining optimal parameter values. In the case of Siamese Networks, theloss is a similarity measure between the latent vector outputs , enforced by abinary class label (like or unlike). The energy function takes as input the la-tent vectors formed by the CNN’s at their last dense layer (for each pair inputpassed) and outputs an energy value. The overall goal during the training pro-cess (optimization) can now be conveyed in terms of the energy function. Theenergy (output of the energy function) of a like pair is minimised and betweenunlike pairs it is maximised.The typical energy functions used could be anything from a simple euclideannorm to a fairly advanced function such as the contrastive energy function. Atypical example of the contrastive energy is explained in brief.he contrastive energy function takes in two vectors as input and in generalperforms the following computation. L ( W ) = P (cid:88) i =1 L ( W, ( Y, (cid:126)X , (cid:126)X ) i ) (1)Where, L ( W, ( Y, (cid:126)X , (cid:126)X ) i ) = (1 − Y ) L S ( D iW ) + Y L D ( D iW ) (2)D W is the parameterized distance function as mentioned below D W ( (cid:126)X , (cid:126)X ) = (cid:13)(cid:13)(cid:13) G W ( (cid:126)X ) − G W ( (cid:126)X ) (cid:13)(cid:13)(cid:13) (3)Y is the binary class label.L S and L D are functions chosen as per the task.Figure 3: Feature comparison methods represented by a Siamese Neural Net-work architecture Data no matter how clean will have irrelevant features, many of the predictive oranalytic tasks does not rely on all of the features present in raw data. One of thefactors that sets us humans apart from computers is our instinct of contextualrelevance while performing any of our day to day activities. Our brains areadept at such tasks which makes us able to perform complex tasks quite easily.Attention is a deep learning technique designed to mimic this very propertyof our brain. Attention, as the name suggests is a methodology by which aneural network learns to selectively focus on relevant features and ignoring therest. Attention was first introduced in the branch of natural language processingNLP) [21], where it enabled contextual understanding for Sequence to Sequencemodels (Ex: Machine Translation) which led to better performance of the same.Attention mechanism in NLP solved the problem of vanishing gradients forRecurrent Neural Networks and at the same time brought in feature relevanceunderstanding which boosts performance. The revolutionary impacts of deeplearning paved way for creation of more efficient network architectures such asthe Transformer (BERT) [22], which are widely applied these days. Moreoverattention has been applied to other fields related to deep learning such as theones focusing on signal and visual processing.
Images are a very abstract from of data, they contain numerous amounts ofpatterns (features) which could be analysed using latest computational tools togain understanding on them. For many machine learning tasks such as regres-sion or classification, identifying features of contextual relavance would improvethe model performance and simplify the task. The same is the case for ma-chine learning applied to images. For most images, the regions could be broadlyclassified as background and objects, where objects are of prime focus and back-ground doesn’t contribute to inference. Because of the same, knowing where tolook and what regions to focus on while making an inference from images helpsboost the performance of the model. Convolutional neural networks(CNN) areone of the best feature extraction tools for images in today’s deep learning lit-erature, attention applied to Convolutional features will help pick out relevantfeatures of interest from the large pool of features extracted by a CNN.Figure 4: A visualization of the attention mechanism applied to Convolutionalfeatures overlaid on X-ray images
Related Work
The outbreak of the COVID-19 [23] pandemic and the increasing count of thenumber of deaths have captured the attention of most researchers across theworld. Several works have been published which aim to either study this virusor in a way aim to curb the spread. Owing to the supremacy of computervision and deep learning in the field of medical imaging, most of the researchersare using these tools as means to diagnose COVID-19. Chest X-ray (CXR) andComputed Tomography (CT) are the imaging techniques that play an importantrole in the detection of COVID-19 [24], [25].As inferred from literature Convolutional Neural Network (CNN) remainsthe preferred choice of researchers for tackling COVID-19 from digitized im-ages and several reviews have been carried out to highlight it’s recent contri-butions to COVID-19 detection[26]-[28]. For example in [29] a CNN based onInception network was applied to detect COVID-19 disease within computedtomography (CT). They achieved a total accuracy of 89.5% with specificity of0.88 and sensitivity of 0.87 on their internal validation and a total accuracyof 79.3% with specificity of 0.83 and sensitivity of 0.67 on the external testingdataset. In [30] a modified version of the ResNet-50 pre-trained network wasused to classify CT images into three classes: healthy, COVID-19 and bacterialpneumonia. Their model results showed that the architecture could accuratelyidentify the COVID-19 patients from others with an AUC of 0.99 and sensitivityof 0.93. Also their model could discriminate COVID-19 infected patients andbacteria pneumonia-infected patients with an AUC of 0.95, recall (sensitivity)of 0.96. In [31] a CNN architecture called COVID-Net based on transfer learn-ing was applied to classify the Chest X- ray (CXR) images into four classesof normal, bacterial infection, non-COVID and COVID-19 viral infection. Thearchitecture attained a best accuracy of 93.3% on their test dataset. In [32]the authors proposed a deep learning model with 4 convolutional layers and 2dense layers in addition to classical image augmentation and achieved 93.73%testing accuracy. In [33] the authors presented a transfer learning method witha deep residual network for pediatric pneumonia diagnosis. The authors pro-posed a deep learning model with 49 convolutional layers and 2 dense layersand achieved 96.70% testing accuracy. In [9] the authors proposed a modifiedCNN based on class decomposition, termed as Decompose Transfer Composemodel to improve the performance of pre-trained models on the detection ofCOVID-19 cases from chest x-ray images. ImageNet pre-trained ResNet modelwas used for transfer-learning and they also used Data Augmentation and his-togram modification technique to enhance contrast of each image. Their pro-posed DeTraC-ResNet 18 model achieved an accuracy of 95.12%. In [34] theauthors proposed a pneumonia chest x-ray detection based on generative ad-versarial networks (GAN) with a fine-tuned deep transfer learning for a limiteddataset. The authors chose AlexNet, GoogLeNet, Squeeznet, and Resnet18 areselected as deep transfer learning models. The distinctive observation drawn forthis paper was the use of GAN for generating similar examples of the datasetbesides tackling the problem of overfitting. Their work used 10% percent ofhe original dataset while generating the other 90% using GAN. In [35] theauthors presented a method to generate synthetic chest X-ray (CXR) imagesby developing an Auxiliary Classier Generative Adversarial Network (ACGAN)based model. Utilizing three publicly available datasets of IEEE Covid ChestX-ray dataset[10], COVID-19 Radiography Database [36] and COVID-19 ChestX-ray Dataset [37] the authors demonstrated that synthetic images producedby the ACGAN based model could improve the performance of CNN(VGG-16in their case) for COVID-19 detection. The classification results showed an ac-curacy of 85% with the CNN alone, and with the addition of the syntheticallygenerated images via ACGAN the accuracy increased to 95% . Thus havingunderstood the advantages that GAN offers on training models with relativelysmaller datasets, in our research we implemented the DAGAN combined withthe attention based Siamese Neural Networks for getting the optimum resultsout of a relatively smaller dataset used for training our model[10].
For our experiments the application was build using Android studio. MVVM(Model View View-Model) architecture has been used in the app, which helpsin proper state management following the UI Material Design guidelines whilebuilding. For storage of the App data(like the local X-Rays samples) and Au-thentication, Firebase is used. The deep learning model was trained using pub-licly available datasets to test for robustness of the same. Some of the majorissues with such datasets were lack of data, inherent noise features and classimbalance. In the proposed methodology, all three of these issues were tackledeffectively. The smartphone application would pave way for improvement of theexisting model and provide ease of access to state of the art disease diagnosis forcommon pulmonary diseases to everyone.A semi-live training scenario was buildon the cloud which enables the model to imporve over time gradually withoutintervention.
The android application acts as an accessible platform which assists doctors orpatients in uploading the X-rays samples to be inferred by the deep learningmodel, and obtain corresponding diagnosis results, as seen in Figure 5. Theapplication is as a cloud-user interface which enables wider accessibility andhelp the model imporve by the cloud build semi-live training scenario. Thealgorithm is deployed in a FAS ( Function as a Service ), which gets triggeredwhen a user uploads a sample. There are primarily 2 categories of users : doctorsand patients. Users under the doctor category would be verified and could actas a potential source for labeled training data. Under the patient category theinference mechanism gets triggered which enables the backend model to providethe user with a diagnosis result for the uploaded sample.igure 5: Abstract workflow of our smartphone application utilizing an AI cloudplatform
The backend deep learning architecture mainly consists of two deep learningmodels- the DAGAN for robust and effective data augmentation, followed bythe Convolutional Siamese Network with attention mechanism. The SiameseNetwork is proven to be data efficient through our experiments. Both of thesenetworks are pretrained on publicly available datasets. To obtain the pretrainedDAGAN model , suitably processed X-ray images were provided with corre-sponding class labels. For the pretrained Siamese Network, visually variantaugmented samples with in-class features preserved were generated using theDAGAN model. Then these generated samples were paired up for all possiblecombinations. Each of the pairs were assigned a binary label based on the classeson which the two images in a pair belonged to - 0 if both images are from thesame class of pulmonary diseases and 1 otherwise. The resulting dataset wasthen used to train the Siamese Network. A set of well labelled and noise freeimages are selected to be the standard dataset for comparison. During inferenceprocedure one of the twin among the Siamese Network generates a latent vectorfor the uploaded image by a forward pass. The second twin generates a latentector for an image in the standard dataset. The obtained latent vectors arethen compared using an energy function. The energy values of all classes in thestandard dataset are obtained using a similar procedure, and the class with thelowest average value is selected. The class thus selected becomes the diagnosisfor the particular uploaded image. The diagnosis made is conveyed back to theuser through an online database.Figure 6: General inference process for images .3 Semi-live training
X-Ray images used by the backend model could show large variance due to avariety of reasons, which includes lighting condition while the picture is taken,the X-Ray machine specifications or camera quality of the user’s smartphoneetc. Since the challenges such as this due to data variation should be accountedin a real world scenario, a semi-live training scenario was introduced, whichenables the model parameters of the pretrained model to further adapt to newor variant data. The scenario is triggered when sufficient amounts of data isobtained.Figure 7: Overview of the training process of the Siamese Network during semi-live training scenario
Experiments and Results
Two different datasets are used to obtain a variety of comparison results forproper model evaluation. On the first and second datasets the tasks are formu-lated as binary and multiclass classification respectively. Dataset descriptionand corresponding comparison results are given below. For the selection of thestandard dataset, expert advice and third party help were utilized.
Figure 8: Class distribution of dataset - 1 [10] (Graphical representation)Figure 9: Class distribution of dataset - 2 [38] (Graphical representation)oth the datasets used in this study are publicly available. Apart from theselection of the standard dataset, no specific dataset cleansing was done. Train-ing process was done on the data including even those images with inherent noisefeatures present, the same helps in confirming the robustness of the proposedmodel. Dataset-1 was published in 2018, the X-rays images obtained were partof clinical care conducted year to year from Guangzhou Medical Center from5,863 different patients. Dataset-2 was published as an effort to give out rele-vant data for widespread studies that were conducted to tackle the COVID-19pandemic situation. Test set size of datasets 1 and 2 were selected as 20% ofimages from each class. The training set was further enlarged using the DAGANmodel to ensure a generalized training for the proposed model.
A good amount of images were selected as the testing set for the proposed model,so as to robustly test and evaluate the proposed method. As per the split of20%, data from each class of the two datasets were randomly selected to beincluded in the test set. For dataset(1) used for the binary classification taskthe test set consisted of 1170 images out of the 5860 images, as for dataset(2)used for the multiclass classification task the test set consisted of 180 imagesout of the 905 images. No generation of images were done on the testing set, asit is considered important to conduct model evaluation on real world data sam-ples.Since the testing set in both experiments are large, the confidence intervalfor the testing accuracy of the proposed model was calculated by assuming aGaussian distribution for the dataset proportion.
Method Year Description TestingAccuracy [39] 2018 Convolutional Neural Network (CNN) 92.80%[40] 2019 Deep learning model with 4 convolutionallayers and 2 dense layers + classicalAugmentation 93.73%[41] 2019 Deep learning model with 7 convolutionallayers and 3 dense layers 95.30%[42] 2019 Deep learning model with 49 convolutionallayers and 2 dense layers 96.70%[43] 2020 Convolutional Neural Network (CNN)+ Random forest 97.00%[34] 2020 GAN + Resnet18 99.00%ProposedMethod 2020 DAGAN + Attention Siamese Net 99.30 +/- 0.63%Table 1: Comparison of testing accuracy of proposed model with related worksconducted on dataset-1 ethod Year Description TestingAccuracy [44] 2020 Using pre-trained model of CheXNet 90.50%[45] 2020 Extracts the features from chest x-rayimages using FrMEMs moment 96.09%[46] 2020 Two-level Hierarchical Deep NeuralNetwork and transfer learning 97.80%ProposedMethod 2020 DAGAN + Attention Siamese Net 98.40 +/- 2.18%Table 2: Comparison of testing accuracy of proposed model with related worksconducted on dataset-2The tables illustrate the robustness of the proposed model as well as pointout how effective the model is under the scenario, where there is a lack ofavailability of training data.
For our purposes, in order obtain a robust and deployable model, we combineboth datasets and train a multiclass classification model which is robustly eval-uated for performance. The testing set is selected to be 20% of images fromeach class, at random. The model thus obtained, achieved a testing accuracyof 97.8%. The validation set is selected to be 20% of images in each class, fromthe training set. Figure 10: The validation loss vs epoch curveDetailed analysis was carried to determine the class separation boundary ofthe latent space, Figure 11 illustrates the large class separation found.igure 11: Range of variation of dissimilarity index for like and unlike classesThe illustration (Figures [12-14]) shows how effective is the latent spacerepresentation so formed by training the model, in representing the lower di-mensional projection of images.Figure 12: Comparing a chest X-Ray image of a COVID-19 positive patientwith test set images in the COVID-19 classigure 13: Comparing a chest X-Ray image of a COVID-19 positive patientwith test set images in the Pneumocystis classFigure 14: Comparing a chest X-Ray image of a COVID-19 positive patientwith test set images in the Normal classThe dissimilarity index values are high when unlike classes are compared, atthe same time a low dissimilarity is obtained when like classes are compared. onclusion
In the wake of the global pandemic preventive and therapeutic solutions are inthe limelight as doctors and healthcare professionals work to tackle the threat,with diagnostic methods having extensive capabilities being the need of thehour. The COVID-19 outbreak has caused an adverse effect in all walks of dayto day life worldwide. Fact remains that the spread of such a disease could havebeen prevented during the early stages, with the help of accurate methods ofdiagnosis. Medical images such as X-rays and CT scans are of great use when itcomes to disease diagnosis, particularly chest X-rays being pivotal in diagnosisof many common and dangerous respiratory diseases. Radiologists can infermany crucial facts from a chest X-ray which can be put to use in diagnosingseveral diseases. Today’s AI methods that mimic disease diagnosis as doneby radiologists could outperform any human radiologist, owing to the higherpattern recognition capabilities and the lack of the human element of error orinefficiency in turn paving way for extensive research in this area. A commonand efficient method to employ would be to use a Convolutional Neural Network(CNN) based classifier , which could accurately recognise patterns from imagesto make necessary predictions. The limitation of the same being the requirementof huge amounts of data to obtain a classifier model with enough generalizabilityand accuracy. Metrics improvement of an existing model is a hard task sinceretraining process for a large deep learning model would be expensive in termsof time and computation, vulnerable to scalability issues in retraining. Hencewe adopted feature comparison based methods which are superior to featurerecognising methods in these respects, exploiting a deep Generative Network fordata augmentation. The model exhibited profound comparison metrics havingvery distinguishable dissimilarity indices. Similar classes showed remarkablylow indices ranging from 0.05 to 0.6 , while different classes had higher valueslying between 1.98 and 2.67. These performance indices of dissimilarity and thethe large gap between these classes consolidates the fact that our model is ableto clearly demarcate and classify diseases with state of the art efficiency.The limitations of this study include the inevitable noise factor on the datasetused, to tackle the same a cloud based live training method has been employedwhich uses properly annotated and identified data from medical practitionersworldwide. The underlying method could be employed to detect several otherdiseases if necessary, modification required for the current model minimal ascompared to any deep learning based backend systems. Doctors and Radi-ologists can leverage the ability of our application to make a reliable remotediagnosis, thereby saving considerable time which can be devoted to medicationor prescriptive measures. Due to the high generalisability and data efficiency ofthe method , the application could prove itself to be a great tool in not only inaccurately diagnosing diseases of interest, but to also conduct crucial studies onemerging or rare respiratory conditions. ode Availability
The custom Python code and android app used in this study are available fromthe corresponding author upon reasonable request and is to be used only foreducational and research purposes.