A new AI assisted medical molecular image diagnostic model

Abstract

In recent times, ML algorithms that plays a significant role right from drug discovery to clinical decision making. The recent advances in DL technologies contribute towards improved performance for carrying out computer aided medical image analysis and disease diagnosis. The key benefit of AI in processing of medical big data offers spectacular insights into the hierarchal relationships that exist among data which can be algorithmically explored thus replacing the tedious manual processes to extract and localize specific areas of interests in medical images thus considerably changing the way medicine has been practiced so far. In bio medical related clinical applications, there is a constant demand pertaining the research and development with respect to deploying AI as a mainstream tool to perform several medical imaging activities like analysis, diagnosis, segmentation as well as classification. The increased usage of electronic health records and medical images being its integral component the need for appropriate and efficient AI assisted medical image analysis system that takes care of accurate and automated decision making could be of great help to radiologists and medical practitioners. Molecular image analysis is a dynamic field that makes use of ML and DL algorithms that utilizes labeled and structured information which also proves to be helpful to the patients as they serve as an initial interface before further diagnosis and treatments. Thus our research aims to offer a novel and efficient AI based medical analysis system that can assist clinical practitioners to focus on enhancing the disease diagnosis through DL based medical image analysis and decision making. In addition, we also address specific challenges related to disease diagnosis and propose novel GAN model for improved diagnosis and implementation. Our proposed technique can also be generalized to generate synthetic data for further issues related to molecular image analysis in the field of medicine and help towards building a better disease diagnosis model.

Keywords

Artificial Intelligence (AI)Deep learning (DL)electronic health records (EHR)Generative Adversarial networks (GAN)medical image analysis Machine learning (ML).

Abbreviations used

Machine Learning

Soft Computing

Genetic Algorithm

Artificial Intelligence

Deep Learning

Reinforcement Learning

Contrast Agent

ANN

Artificial Neural Network

GAN

Generative Adversarial Networks

SVM

Support Vector Machines

CNN

Convolutional Neural Networks

RNN

Recurrent Neural Networks

MSE

Mean Square Error

Fuzzy Logic

Computer Tomography

PET

Positron Emission Tomography

ECG

Echo cardiogram

PSNR

Peak Signal To Noise Ratio

EHR

Electronic Health Records

1 Introduction

Digital advancements like AI is transforming and revolutionizing nearly every business known to mankind, yet there exists one industry that stands to gain the most from this technology which is health care. This is especially due to the fact that integrating human intelligence with AI technology is likely to benefit every stakeholder [4] involved in medicine and pharmaceutical industries right from patients, radiologists, drug store personnel to research experts and clinical practitioners (Fig. 1).

Fig. 1

Applications of AI in Healthcare industry.

Recent trends in medical technology along with necessary funding has been witnessing several life saving health care facilities that utilizes advanced machineries that perform multiple functions like medical data storage, analysis and diagnosis to gene sequencing operations.

Arrival of Big data has further resulted in unprecedented discoveries that were beyond natural human comprehensive abilities when it comes to interpretation and analyzing the tremendous medical data acquired from multiple resources available. Wide range of bio-sensors and machine learning algorithms facilitates the storage as well as uninterrupted measurement of patients’ fitness parameters. The transformation of AI into implementable biomedical and information has resulted in precision medicine where ML algorithms portrays the potential to learn and thus progress over time. The key benefits of deploying these AI algorithms or DL based neural networks in the field of medicine has already generated remarkable changes in health care applications where lesser diagnostic errors, cost effective utilization of resources, and letting the medical practitioners to tend to the needs of other patients. Before implementing AI algorithms in biomedical establishments, multiple challenges need to be resolved to accomplish the desired outcome. Machine learning is mostly data driven where machines acquire the ability to learn through training and inference. Training lets the machine to discover and explore the presence of patterns and existence of relationships that exists among the data in comparison with previously acquired data, whereas inference evaluates these patterns to perform operations like prediction and decision making. ML algorithms have been constantly evolving and improvising [5 , 25], thus becoming more and more accurate and sophisticated giving rise to advanced hierarchical structures called Deep Learning. DL has the unique capability of getting organized in a hierarchical fashion consisting of multiple levels thus automatically extracting significant features from the input data. A DNN consists of an input layer, several hidden layers and an output layer and setting up of these layers is dependent on the input data. Well-known patterns are employed to train these networks to achieve grouping of definite inputs together or individually. These neural networks have presented significant computational achievements yet suffer from certain constraints especially with respect to limited availability of data to perform training procedures [12]. Another aspect being available data must be unbiased and the big data that AI relies may not always be valid all the time. To evade these issues, Generative Adversarial Networks (GAN) can be utilized to produce synthetic datasets (Fig. 2) which are needed to train deep neural networks [11].

Fig. 2

Architecture of GAN.

The GAN model architecture has 2 components namely: a generator that produces new instances and a discriminator that performs classification to validate whether generated instances are real or fake. Thus GAN is based on game theory where generator must compete with potential adversary where generator creates samples whereas adversary acts as a discriminator and distinguishes between instances drawn from training data and that of instances generated by generator [1 , 8].

The extensive availability of multi-level biomedical data has offered us with the opportunity to learn the nature and existence of several biological operations. In comparison to traditional data investigation [21], AI is steadily exerting an ever increasing effect in the bio medical field where processing algorithms and necessary hardware facilitates development of model infrastructure in parallel with learning process. There is always a demand for AI-assisted biomedical applications that can efficiently analyze, classify, diagnose and predict future ailments [25]. Unlike conventional image and signal processing approaches, the AI algorithms help out the doctors by enhancing their intelligence and diagnosis prediction accuracy and classification as they have the ability to automatically retrieve and analyze the features without getting affected by prejudiced aspects. Several ML algorithms like SVM, CNN, FL, and RNN are utilized in medical domain which is appropriate for diverse application requirements. For instance, SVM is employed for processing ECG signals whereas fuzzy logic based classifiers are widely used in monitoring vital signs like blood pressure and glucose. CNNs can outperform human beings in medical domains especially diagnosing from medical visual data. In bio medical applications exclusive hardware’s are used to provide dedicated services for AI based biomedical processing procedures through lesser consumption of energy and resources. Software platforms are required to perform functionalities like disease detection, health monitoring etc., through their architectures that integrates image processing modules to carry out filtering, de-noising, segmentation and extraction of features along with dimensionality reduction and classification to carry out accurate diagnosis [11, 21]. Thus our research is focused on designing an optimized architecture for AI assisted biomedical practices, with the objective of achieving highest possible medical images classification accuracy along with maintaining the reduced computational complexity.

Contrast media used in radiology is mostly iodine-based which is essential for performing several diagnostic functions as the growing demand for PET and CT scans persists to rise globally. The major limitation in using these contrast agents is the unavoidable exposure of patients to these agents whose side effects ranges from allergies and hormone imbalances to renal failure. Radiologists worldwide are increasingly facing with the challenge of arriving at an accurate diagnostic administration of these intravenous agents to prevent any complications that may result in hospitalization, medical issues or even mortality. Optimizing the utility of CA through technical and enhanced reconstruction procedures through the recent advancements of AI where medical image post-processing can further prevent many limitations due to overexposure of CA thus saving multiple resources and cost. Thus our proposed research aims to offer the following contributions namely

Construction of AI assisted medical diagnostic model using Generative Adversarial Networks that reduces the amount of contrast agents like iodine intake thereby minimizing many of the side effects that occurs due to their high dosage.

Design and development of an optimized framework that assists the medical practitioners and clinical personnel towards arriving at informed decisions through effective utilization of time and resources by generating high quality medical images.

Enhancing the competence of radiologists and pathologists through cost–effective and time saving workflow.

Integration of AI and human intelligence to accommodate existing limitations and complement human resources by generation of synthetic data sets.

The organization of this study is as follows: an introduction on role of AI in healthcare along with an overview of GAN together with issues and challenges encountered by traditional approaches in medical care has been presented in the first section.The second section explores on multiple literatures in relation to our specific research and offers an insightful review on multiple and up to date works in the field of AI and medicine. The third section deals with proposed system model detailing the architectural framework utilized to implement the AI assisted medical imaging system. The results are analyzed and discussed in section 4 followed by conclusion which concludes the research by stating the significant details in section 5 along with possible future scopes and directions.

2 Literature survey

Medical imaging is an important practice for the analysis and diagnosis of diseases in healthcare management. Soft Computing (SC) has a special role in modern advancements in medical imaging scenarios where the image abnormalities are enhanced and quality of the images is improved to facilitate accurate diagnosis and treatment. Multiple SC techniques like FL, GA, ANN, CNN are utilized for healthcare applications [14]. A detailed review of various SC technologies has been presented in detail followed by a comparative analysis on various metrics like accuracy, specificity, noise reduction, error rates etc., Mor et al [16] demonstrated an innovative approach to identify and help children with a specific learning disorder (SLD).The identification is done using Deep CNN which helps in rapid and efficient screening of kids through application of certain criteria to check for any possible learning disorder and thus helps in early intervention.

Koshino et al [10] presented an extensive review on application of GAN to facilitate generation of synthetic images based on ANN and DL. Along with providing adequate reliability and adaptability inbuilt in DL, GANs offer the potential ability to aid problem-solving that has garnered attention and is being applied for medical and molecular imaging applications. The author provides a complete overview of GANs and talks about its effectiveness in medical domain. Weintraub et al [24] insisted on translational medicine in this age of big data and ML which could actually optimize the healthcare by facilitating better patient care at a reduced cost, enhancing the overall health of the population. AI thus augments the capacity of professionals involved in the processes of diagnosis and risk predictions can offer the best treatment option thereby speeding up the workflow as well as accuracy in the medical care scenarios.

Etcheverry et al [15] gave a nonlinear localization based AI methodology to achieve high accuracy in medical image based analysis. This method of locating anatomical arrangements helps to segment the images, quantify and detect the presence of any abnormalities. Deep RL based technology is utilized to assist the learning process which actively locates the object in the input image. In accordance with the object parameterization, the agent discovers and optimizes the parameters through a string of control actions. Esteva et al [1] suggested some studies where DL algorithms can efficiently carry out diagnostic operations at par with human professionals and also present reports where AI algorithms can surpass humans. Yet it’s always better for AI to work in synergy with human intelligence while arriving at life saving and emergency situations.

Iqbal et al [22] proposed a GAN for performing Medical Imaging which generates synthetic medical images along with segmented masks that can be applied for analysis of medical images. Retinal images were used to conduct experiments and the suggested method generated accurate segmented images in comparison to existing methodologies. Park et al [20] insisted on making use of AI as an essential methodology to carry out clinical evaluation in the field of medicine through utilization of high-dimensional diagnostic models in like artificial DNN to determine life threatening conditions that demands immediate attention and requires the action by concerned specialists to perform particular treatment procedure. Ehteshami et al. [3] assessed the preciseness and effectiveness of DL algorithms in diagnosing breast cancer patients in comparison with clinical pathologists. Their results confirm that quite a lot of algorithms were able to outperform the pathologists especially in a time based diagnostic setting. In normal situations, pathologists require sufficient time to evaluate and analyze the slides. But DL algorithms were effective in rapidly analyzing and generating predictions based on the input slides in comparison to lab personnel and were free from errors and other common prejudices.

Development in technologies lead to several non invasive diagnostic methods like monitoring the bodily factors like heartbeat, oxygen levels, temperature, pressure etc., at molecular level where Individuals can continuously monitor and detect minute changes [17]. Any variations from normal can be noticed early, thus saving both lives as well as resources. Murdoch et al presented the unavoidable relation between AI and Big data [21]. The role of AI in several fields like astronomy, social media, medicine, politics, retail etc., in exploring and discovering essential information from these large datasets and to help specialists in respective domains in making accurate decision [9, 13]. However, all these techniques so far discussed above are applicable only for the high resolution image. Hence to overcome this drawback, the proposed model (Fig. 4) developed a new framework to analyze medical images and generate high quality resolution images using Generative Adversarial Networks (GAN) that facilitates reconstruction of trustworthy and realistic visual medical images.

3 GAN model principle

Thus the demand for medical image reconstruction through alternate means has increased as the limitations with contrast agents with respect to higher dosage has been affecting millions worldwide. Owing to its complexity along with maximum visual requirements of medical images, GAN is employed which takes the lesser resolution image as input and generates higher quality medical image as an output. In our proposed system, an enhanced medical image generation network has been employed comprising of an image generator and a discriminator to differentiate between lesser and better quality of input medical images. Loss functions integrating the loss due to content and features are performed to obtain the resultant medical images with better accuracy and reliability [2, 7].

This proposed model has two main components

Generator module

Discriminator module

A training set containing N number of medical image samples is provided. The Generator module takes as input as a random vector and attempts to generate images related to those in the training set. Discriminator module acts as a binary classifier which performs the differentiation between the real medical images as per the input training set samples and the unreal medical images generated by the Generator module (Fig. 3) [18 , 23].

Fig. 3

Architecture of Discriminator module.

Fig. 4

Proposed system architecture utilizing GAN for medical image reconstruction and analysis.

Thus, the task of the Generator lies in learning the data distribution in the training set, so that it can recreate factual looking medical images and ensure that the Discriminator cannot discriminate between those input images from the training set and that of images produced by the Generator module.

The generator module has the following architectural components

Makes use of 3 2D transpose convolution layers.

Every layer has a stride of 4 and kernel size of 6.

Batch normalization is carried out in every single layer.

ReLU is the activation function used.

Generator module maps a 200 vector of noise in the range of [-1,1] to a 4×4×256 tensor to generate 2×2×256 image after operations on all the layers.

The discriminator module has the following architecture Consists of 3 2D transpose convolution layers

Every layer uses a stride of 4 and kernel size is 6.

Batch normalization procedure is done using Leaky ReLU activation.

Logistic regression is applied for the output layer

Let the image generated by generator network with corresponding low resolution original image(OLR) be HQR, thus $HQR = Gen (OLR)$ (1)

Gen() → Generator module operation and QR is the ground truth.

An image pair (OLR, HQR/QR) is given as an input to discriminate the high resolution image (HR) with a given OLR image. With our proposed design, the discriminator module is made to learn this pair-wise information of both HQR/QR and OLR medical images by means of feature concatenation procedure extracted from the OLR and HQR and outputs the probability as (OLR, HQR) or (OLR, QR) pair as an actual pair.

With CT slice pair of (OLR, QR) as 1 and (OLR, HQR) as 0, the training stage for discriminator. Module is expressed as:

$\begin{matrix} Dis (OLR, QR) \to 1 \\ Dis (OLR, HQR) \to 0 \end{matrix}$ (2)

Where Dis() is the discriminator function.

Throughout the training process, either the Generator or the Discriminator module is trained varyingly using dissimilar data during iterations to deal with constraint with respect to model parameters getting trapped into local optimum. Thus loss is calculated through cross entropy to train the Discriminator module

$\begin{matrix} L_{Dis} = - log (Dis (OLR, QR)) \\ - log (1 - Dis (OLR, HQR)) \end{matrix}$ (3)

Where Dis() is the function of Discriminator module.

Thus the total loss encountered by Generator module includes content loss (L_con), adversarialloss (L_adv) as well as feature loss (L_fe) which is estimated as $L_{Gen} = L_{con} + μ_{1} L_{adv} + μ_{2} L_{fe}$ (4) where μ₁ and μ₂ are hyper parameters that can be fine tuned for balancing the effects of each element of loss.The content loss can be expressed as $con = \frac{1}{N^{2}} \sum_{i = 1}^{N} \sum_{j = 1}^{N} | QR (a, b) - HQR (a, b) |$ (5)

This loss is solely accountable for the restoration of contents in medical images, where N represents the size of medical image samples, assuming that every image is N x N. Through optimized minimization of L_con, reconstruction of the high frequency content and thereby preserving the actual visual characteristics of the medical image is possible in realistic manner. As far as medical images are concerned, texture based features are very critical as only these maximum frequency visual data plays a vital role in judgment by human experts who rely on the quality of images to make informed decisions and appropriate treatment further, thus mere estimation of PSNR based loss is not just adequate for medical images based diagnosis. Extraction of texture based features present on hidden layers is essential to calculate the total content loss which fully extracts the features on every semantic level and integrates losses with higher weights on shallower image blocks and minimal weights on much deeper image blocks. This is due to the fact that deeper features generate higher abstract semantic knowledge, while shallow features portrays relatively higher concrete information which can be expressed as

$\begin{matrix} L_{con} = μ_{L 1} \cdot L_{1} (QR, HQR) + \sum_{n = 1}^{5} w_{n} \cdot \\ MSE (k_{{(a, b)}_{n}} (QR), k_{{(a, b)}_{n}} (HQR)) \end{matrix}$ (6) where μ_L1 is a tunable hyper parameter for balancing the direct loss between HQR and QR image, weights w={1/2.1/4,1/8,1/16} and layers are (a,b)={(1,2),(2,3),(3,4),(4,5)} for the hidden blocks, MSE is the mean square error of two features. Feature maps are employed before ReLU activation in generator module for all hidden blocks to evade the information lost occurring after activating batch normalization procedure.

Adversarial loss intends to aid the Generator module to make HQR image closer to realistic QR image as much as possible with the main idea of deceiving the Discriminator module. Thus adversarial loss L_adv is the reciprocal of L_dis. Hence Dis(OLR, HQR) must be made closer to 1: $L_{adv} = - log (1 - Dis (OLR, HQR))$ $- \log (Dis (OLR, QR))$ (7)

Reducing the feature loss through minimizing the differences between texture features helps generator module to get benefited by creating more realistic samples.

With Dis_n(., .) represents the n-th hidden block inside the discriminator module and the feature loss is defined as:

$\begin{matrix} L_{fe} = \sum_{n = 1}^{5} w_{n} \cdot MSE \\ (Di s_{n} (OLR, QR), Di s_{n} (OLR, QR)) \end{matrix}$ (8) Thus the final discriminator function can be updated as

$\begin{matrix} D is (n) \\ = \frac{e^{[- \frac{1}{2} {(L_{con} - μ_{2})}^{n} Σ^{- 1} (L_{con} - μ_{1})]}}{{e^{[- \frac{1}{2} {(L_{adv} - μ_{2})}^{n} Σ^{- 1} (L_{adv} - μ_{1})}]}_{n} e^{[- \frac{1}{2} {(L_{fe} - μ_{2})}^{n} Σ^{- 1} (L_{fe} - μ_{1})]}} \end{matrix}$ (9)

where $\begin{matrix} μ_{1} = \frac{\sum_{i = 1}^{n} a_{1}^{(i)}}{b} \\ \sum = \frac{1}{a + b - 2} [\sum_{i = 1}^{a} (μ_{2}^{{i}} - μ_{1}) {(μ_{2}^{{i}} - μ_{1})}^{n} \\ + \sum_{i = 1}^{b} (μ_{2}^{{i}} - μ_{2}) {(μ_{2}^{{i}} - μ_{2})}^{n}] \end{matrix}$

Pseudo code for proposed GAN model:

Initialize the number of input medical image samples(N) Input: Medical image samples

Output: Reconstructed high quality medical images Start

Estimate the content, adversarial and feature loss for generator through mean square error values for real samples Sample the noise samples

Acquire synthetic medical image samples from generator through noise samples

Update the parameters of discriminator using generated samples and real samples

$\begin{matrix} μ_{1} = \frac{\sum_{i = 1}^{n} a_{1}^{(i)}}{b} \\ \sum = \frac{1}{a + b - 2} [\sum_{i = 1}^{a} (μ_{2}^{{i}} - μ_{1}) {(μ_{2}^{{i}} - μ_{1})}^{n} \\ + \sum_{i = 1}^{b} (μ_{2}^{{i}} - μ_{2}) {(μ_{2}^{{i}} - μ_{2})}^{n}] \end{matrix}$

Attain the final probability function for discriminator module

Update the hyper parameters of generator and discriminator

Terminate the processes

End

4 Results analysis and discussion

The evaluation of our proposed model was carried out on CT and MRI scans and the results portrayed significant influence on disease diagnosis, that reveals that our GAN based medical imaging systems can possibly be applied in practical applications. Datasets include scans from LUNA 16 Challenge, 200 randomly selected scans that were used for training and testing. Our networks take an input of 128×128 CT slice and generates an output a 512×512 CT slice.

Our networks are fully convolutional thus can be fed with arbitrary CT slice sizes and generate fourfold of these slices. To train the generator, we set μ₁ =10^–4, μ₂ = 10^–4 and μL_cont=0 for loss function and weights of content loss is assigned as four orders of magnitude larger than the weight of adversarial and feature losses. Fine tuning of μ₁ and μ₂ during experiments is done but effects were limited on generated medical images. The initial learning rate was set to 10^–4 and reduced during the end of 100 iterations. These experiments were conducted on Titan RTX GPUs. An additional set of 50 thoracic CT scans(10,000 slices) were collected from radiological centers and various artificial noises like salt and pepper noise, and certain random noise were added to generate high quality images with resolution of 256 * 256. The list of hyper parameters used in our experiment has been presented below (Table 1).

Table 1
Hyper parameters and values

Parameter Values

Batch Size 32

Image Size 256×256

Filter size 5x5

Epoch 50

Learning Rate 1×10^- 4

ReLU slope 0.2

Momentum 0.5

Parameter	Values
Batch Size	32
Image Size	256×256
Filter size	5x5
Epoch	50
Learning Rate	1×10^- 4
ReLU slope	0.2
Momentum	0.5

Evaluation indicators used in our experiment are (PSNR) and structural similarity index (SSIM). These two indicators are presented in equations as $\begin{matrix} MSE = \frac{1}{N^{2}} \sum_{i = 1}^{N} \sum_{j = 1}^{N} {(i (x, y) - j (x, y))}^{2} \\ PSNR = 10 \cdot lo g_{10} (\frac{MA X^{2}}{MSE}) \\ SSIM = \frac{(2 μ_{i} μ_{j} + x_{1}) (2 σ_{ij} + x_{2})}{(μ_{i}^{2} + μ_{j}^{2} + x_{1}) (σ_{i} + σ_{j} + x_{2})} \end{matrix}$ where I, J, N are generated medical image, ground truth image, and the size of input image respectively, MAX denotes maximum gray scale value and μ, σ represents mean and variance, σij is the covariance, x1 and x2 are constants arrived with the conventions of SSIM.

We have compared the performance of the proposed method (Tables 2, 3) with several state-of-the-art methods for medical image diagnosis. De-noising results of different methods has been presented in Fig. 5. Further verification of results of different methods using their PSNR and SSIM data were done to evaluate the performance of our proposed methodology. Figure 5 illustrates the efficiency of our scheme by comparing with mechanism devoid of RDB to obtain super resolution images. As summarized in Table 2 and 3, the proposed GAN performs the best which indicates that proposed topology can obtain comparatively the best perceptual quality among these methods.

Table 2

Average PSNR and SSIM de-noised images

	PSNR Mean		SSIM Mean
Methods	JSRT	LIDC	JSRT	LIDC
Noise	15.4436	18.3269	0.1156	0.1387
WGAN	16.3215	20.1685	0.7526	0.7589
CGAN	12.5962	18.3254	0.4421	0.6954
SRGAN	16.2154	21.562	0.7745	0.8056
CGAN+LI	32.5985	34.2659	0.8874	0.8547
Proposed Scheme	33.2514	35.1265	0.9214	0.9365

Table 3

Average PSNR and SSIM on CT scan images

	CT				MRI
Method	LUNA 16 Tests		Clinical CT Scans		MRI Tests
	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
Bi-Cubic	24.456	0.985	25.214	0.965	25.136	0.985
RGAN	31.845	0.992	27.621	0.985	26.995	0.979
RWMAN	31.356	0.957	27.254	0.984	27.074	0.991
ESRGAN	31.914	0.98	27.518	0.956	26.151	0.985
MedSRGAN	32.554	0.988	27.548	0.956	26.174	0.985
Proposed Scheme	32.701	0.995	27.769	0.986	27.459	0.992

Fig. 5

Evaluation of generated image.

Fig. 6

PSNR and SSIM on LUNA 16 Tests.

Fig. 7

PSNR and SSIM on MRI Tests.

Figures 6 and 7 depicts the performance measures of Luna 16 and MRI data set. Table 4 and Fig. 8 show the summary of speed across different methods. As per the observations in a controlled environment our proposed scheme has recorded 27.15 slices per second which is the highest of all methods compared.Storage required by the proposed method is also less (42.35 MB). At the same time, the parameter quantity of the method is low, which in turn reduces the complexity of calculation and effectively improves the real-time performance of network processing.

Table 4

Speed of SR methods

Method	Speed (Slices per sec)	Param Count (Million)	Param Storage(MB)
Bi-Cubic	600.95	–	–
RGAN	11.59	15.58	62.63
ESRGAN	27.12	11.79	47.52
MedSRGAN	26.48	10.61	42.61
MedSRGAN+	25.59	10.63	42.62
Proposed Scheme	27.15	10.59	42.35

Fig. 8

Comparison of Speed and other Vital attributes.

Figure 9 depicts the Visual comparison of MRI and CT images from the LUNA dataset.

Fig. 9

Visual comparison of MRI and CT case from the LUNA dataset.

5 Conclusion and future enhancements

Modern years have observed the speedy expansion in the use of AI and ML in medical imaging technology. This all time data-hungry technology requires huge volumes of datasets which are indispensable to meet up the ever growing demand of high-quality super resolution medical images which act as training sets. In order to address multiple disciplines of radiology, be it interventional, neuro sciences or molecular imaging extensive imaging datasets must be aggregated from multiple clinical centers that has to be reviewed, and suitably interpreted by expert professionals for further decisions. Novel approaches are required to overcome the limitations posed by traditional methodologies. GAN architectures acts as a promising AI tool for medical and molecular imaging that performs image synthesis, reconstruction, segmentation and modality conversion to produce super quality images through low-dose agent administration and limited acquisition time. In the present study, GAN based diagnostic model has been utilized to augment the medical images that contrasts selectively and the results ensures reduced contrast media dose and has achieved the better performances in further quantitative analysis where noise ratio has been reduced to greater extent. Lastly, we can conclude that our proposed method eradicates the noise levels in input image samples successfully and generates synthetic high quality samples while preserving actual structural and contrast data of the input images, and thus our proposed method looks promising for practical clinical applications.

Funding statement

The authors received no specific funding for this study.

Conflicts of interest

The authors declare that they have no conflicts of interest to report regarding the present study.

References

Esteva

, Kuprel

, Novoa

R.A.

, Ko

, Swetter

S.M.

, Blau

H.M.

, Thrun

Dermatologist-level classification of skin cancer with deep neural networks, Nature 542 (2017), 115–118.

Agisilaos Chartsias , Thomas Joyce , Rohan Dharmakumar , Sotirios , Tsaftaris

Adversarial image synthesis for unpaired multi-modal cardiac data. In International Work shop on Simulation and Synthesis in Medical Imaging, pages 3–13. Springer, (2017).

Ehteshami Bejnordi

, Mullooly

, Pfeiffer

R.M.

, Fan

, Vacek

P.M.

, Weaver

D.L.

, Sherman

M.E.

Using deep convolutional neural networks to identify and classify tumor-associated stroma in diagnostic breast biopsies, Modern Pathology 31(10) (2018), 1502–1512.

Dong

, Loy

C.C.

, He

, Tang

Learning a Deep Convolutional Network for Image Super-Resolution, Fleet D., Pajdla T., Schiele B., Tuytelaars T. (eds) Computer Vision –ECCV (2014), pp. 184–199.

Kermany

D.S.

, Goldbaum

, Cai

, Valentim

C.C.S.

, Liang

, Baxter

S.L.

, McKeown

, Yang

, Wu

, Yan

, Dong

Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning, Cell 172(5) (2018), 1122–1131.

Greenspan

Super-resolution in medical imagings, Comput J 52(1) (2009), 43–63.

Goodfellow

, Pouget-Abadie

, Mirza

, Xu

, Warde-Farley

, Ozair

, Courville

, Bengio

Generative Adversarial. Networks, ArXiv e-prints (2014).

Sutskever

, Vinyals

, Le

Q.V.

, Ghahramani

, Welling

, Cortes

, Lawrence

N.D.

, Weinberger

K.Q.

Advances in Neural Information Processing Systems 27, Curran Associates (2014), pp. 2672–2680.

Grace

, Salvatier

, Dafoe

, Zhang

, Evans

WhenWill AI Exceed Human Performance? Evidence from AI Experts. arXiv, arXiv:1705.08807v2. 2017.

10.

Koshino

, Werner

R.A.

, Pomper

M.G.

, Bundschuh

R.A.

, Toriumi

, Higuchi

, Rowe

S.P.

Narrative review of generative adversarial networks in medical and molecular imaging, Annals of Translational Medicine 9(9) (2021).

11.

Gondara

Medical Image Denoising Using Convolutional Denoising Autoencoders, IEEE 16th International Conference on Data Mining Workshops (ICDMW) (2016), pp. 241–246.

12.

Gatys

L.A.

, Ecker

A.S.

, Bethge

Image style transfer using convolutional neural networks, IEEE conference on computer vision and pattern recognition (CVPR) (2016), pp. 2414–2423.

13.

Liang-Chieh Chen , George Papandreou , Iasonas Kokkinos , Kevin Murphy , Alan Yuille

Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs, IEEE Transactions on Pattern Analysis and Machine Intelligence 40 (2018), 834–848.

14.

Devi

, Singh

, Tiwari

, Chandra Patel

, Ayana

M.T.

A survey of soft computing approaches in biomedical imaging, Journal of Healthcare Engineering (2021).

15.

Etcheverry

, Georgescu

, Odry

, Re

T.J.

, Kaushik

, Geiger

, Mariappan

, Grbic

, Comaniciu

Nonlinear adaptively learned optimization for object localization in 3D medical images, In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support (pp. 254–262). Springer, Cham.

16.

Mor

N.S.

, Dardeck

Applying a Convolutional Neural Network to Screen for Specific Learning Disorder, Learning Disabilities: A Contemporary Journal 19(2) (2021), 161–169.

17.

Belmonte-Fernández

Ó.

, Puertas-Cabedo

, Torres-Sospedra

, Montoliu-Colás

, Trilles-Oliver

An indoor positioning system based on wearables forambient-assisted living, Sensors 17(1) (2016), 36.

18.

Isola

, Zhu

, Zhou

, Efros

A.A.

Image-to-Image Translation with Conditional Adversarial Networks, In: IEEE conference on computer vision and pattern recognition (CVPR) (2017), 5967–5976.

19.

Yan

, Xu

, Rastinehad

A.R.

, Wood

B.J.

Adversarial Image Registration with Application for MR and TRUS Image Fusion, In: Shi Y., Suk HI., Liu M. (eds) Machine Learning in Medical Imaging, Lecture Notes in Computer Science, l 11046. Springer, Cham (2018).

20.

Park

S.H.

, Han

Methodologic guide for evaluating clinical performance and effect of artificial intelligence technology for medical diagnosis and prediction, Radiology 286(3) (2018), 800–809.

21.

Murdoch

T.B.

, Detsky

A.S.

The inevitable application of big data to health care, Jama 309(13) (2013), 1351–1352.

22.

Iqbal

, Ali

Generative adversarial network for medical images (MI-GAN), Journal of Medical Systems 42(11) (2018), 1–11.

23.

Tero Karras , Timo Aila , Samuli Laine , Jaakko Lehtinen Progressive growing of gans for improved quality, stability and variation, arXiv preprint arXiv:1710.10196, (2017).

24.

Weintraub

W.S.

, Fahed

A.C.

, Rumsfeld

J.S.

Translational medicine in the era of big data and machine learning, Circulation Research 123(11) (2018), 1202–1204.

25.

Guo

, Cheng

, Nie

, Wang

, Ma

, Kankanhalli

Attentive long short-term preference modeling for personalized product search, ACM Transactions on Information Systems 37(2) (2019), 1–27.

A new AI assisted medical molecular image diagnostic model

Abstract

Keywords

Abbreviations used

1 Introduction

3 GAN model principle

Table 1 Hyper parameters and values Parameter Values Batch Size 32 Image Size 256×256 Filter size 5x5 Epoch 50 Learning Rate 1×10- 4 ReLU slope 0.2 Momentum 0.5

Funding statement

Conflicts of interest

References

Table 1
Hyper parameters and values

Parameter Values

Batch Size 32

Image Size 256×256

Filter size 5x5

Epoch 50

Learning Rate 1×10^- 4

ReLU slope 0.2

Momentum 0.5