AI-based solutions for the analysis of biomedical images and signals

Abstract

Artificial intelligence (AI) has emerged as a disruptive technology that is transforming the medical field. The use of AI algorithms, machine learning, and neural networks has revealed a great potential in enhancing healthcare delivery, including early detection and diagnosis of diseases, improving patient outcomes, and reducing healthcare costs. The application of AI in medical imaging analysis is going to obtaining remarkable success in identifying tumors and other anomalies, leading to earlier diagnoses and better treatment outcomes. In addition, AI has also been applied to medical data analysis, drug discovery, and personalized medicine. In this review article we present an overview of the research work conducted by our team in the field of medical AI. Our primary focus has been on the development and validation of AI-based approaches for the early detection and accurate diagnosis of breast cancer, skin melanoma, and heart disease. Our research has demonstrated the potential of AI in improving the accuracy and eficiency of medical diagnosis and treatment. With the growing availability of medical data and advances in AI technology, we anticipate that the future of medical AI will bring about a paradigm shift in the way healthcare is delivered.

Keywords

• eol>Artificial Intelligence
• neural network
• classification
• data augmentation
• digital breast tomosynthesis
• Generative Adversarial Network
• Deep Convolutional Neural Network
• ECG
• pattern detection
• skin melanoma

1. Introduction

Artificial Intelligence (AI) is emerging as a gamechanging technology in the field of medicine, capable of ushering in a new era of patient-centered care. By leveraging machine learning (ML), deep learning (DL), natural language processing (NLP), and other AI-based technologies, medical practitioners can make more informed and accurate decisions, leading to improved healthcare outcomes. AI-based algorithms can mine vast amounts of data from electronic health records, medical images, genomic data, and other sources, providing clinicians with critical insights and enabling them to make more personalized treatment decisions. Moreover, the potential of AI extends beyond clinical settings, with applications in telemedicine, remote patient monitoring, and public health management. AI-powered devices such as wearables and sensors can help monitor patient health and alert clinicians to potential health risks before they escalate. The development of innovative solutions for early disease diagnosis is a critical objective in the biomedical field. In this context, the project utilizing AI-based techniques to analyze biomedical images and signals is well-aligned. By leveraging advanced algorithms and machine learning models, this project aims to provide accurate and eficient diagnosis results to healthcare professionals. With the potential to detect diseases at their earliest stages, these solutions have the potential to revolutionize the healthcare industry and improve patient outcomes. Overall, our project represents a significant step forward in the quest to develop innovative solutions for the early diagnosis of diseases.

Currently, we are investigating a variety of medical and afect people of all skin types. In this scenario, early imaging techniques, including tomosynthesis, electrocar- detection is critical to ensure patient survival [ 1 ]. diogram (ECG), dermoscopy, and clinical. By leveraging In the following we describe the main results of our the power of AI, we aim at developing more accurate and research related to the melanoma detection. eficient methods for identifying and diagnosing a wide range of conditions, from cardiovascular disease to skin 3.1.1. Minimization of False Negative Rate cancer.

Concerning tomosynthesis analysis, we focused on It is important to provide diagnostic support tools able to Digital Breast Tomosynthesis (DBT) images facing two achieve both high accuracy and minimize type 2 errors, crucial and strongly correlated aspects of deploying Com- related to the False Negative Rate (FNR). Consequently, we puter Aided Diagnosis (CAD) systems: classification and explored the behavior of nine types of CNNs (including data augmentation. The classification task was carried Alexnet, DenseNet, GoogleNet Inception V3, GoogleNet, out using a Deep Convolutional Neural Network (DCNN) MobileNet, ShufleNet, SqueezeNet, and VGG16) on MEDfor mass- and micro-calcification detection, a state-of- NODE, a dataset of 170 clinical images [ 2 ]. To investithe-art image analysis model that was contextually ad- gate the impact of data augmentation and image prejusted. The need for huge volume of data for training processing on the final classification performance, four such models represents a crucial problem due to privacy datasets were generated from the original (INA, NIA, restrictions, great efort for labelling images, costs of the IA, NINA), with pre-processing quality step (IIQ) and medical exams and so on. Thus, there is the need of in- a simple segmentation process (OTSU). INA contains vestigating afordable data augmentation technique. A MED-NODE original images by using quality improved novel approach to generative models, namely Evolution- and data augmentation techniques; NIA is composed by ary Generative Adversarial Network (E-GAN) has been MED-NODE original images not quality improved but usused to perform data augmentation on the same kind of ing data augmentation techniques; IA includes the NSA images. images by using quality improved and data augmenta

Similarly, in the case of ECG, we are exploring how tion techniques while NINA is the original MED-NODE AI can be used to analyze cardiac signals and identify dataset not improved and not data augmented. All tested patterns that may indicate the presence of underlying neural networks perform better without data augmentaconditions. By developing algorithms that can detect tion, with AlexNet and SqueezeNet achieving a maximum these patterns quickly and accurately, we hope to im- accuracy of 78%. Without pre-processing and data augprove the accuracy of diagnoses and ultimately improve mentation, AlexNet performed best with 89%, 75% and patient outcomes. 82% of accuracy, sensitivity, and specificity, respectively,

Finally, in the case of dermoscopic and clinical images, as shwon in Fig. 1. VGG can guarantee the lowest FNR we are exploring how AI can be used to analyze images of at the expense of global accuracy, whereas AlexNet can the skin and detect early signs of skin cancer. By identify- guarantee comparable FNR to VGG but with the highest ing suspicious lesions and providing early intervention, global accuracy. we hope to improve the prognosis for patients with this deadly disease.

These activities involve the University of Salerno, Parthenope University of Naples, and the University of Naples Federico II. The CAISLab laboratory (Computer Science Department of the University of Salerno) provides support for the development of time and costconsuming algorithms.

3. AI Applications on biomedical images and signals

3.1. Classification of melanoma images

Melanoma is a severe form of skin cancer that accounts

for approximately 99,780 new malignant diagnoses each year 1.It come in a variety of shapes, sizes, and colors,

1https://www.cancer.org/cancer/melanoma-skin

cancer/about/key-statistics.html 3.1.2. CNNs design employing genetic algorithms

In [3] we adopted genetic algorithms to designing an architecture for a convolutional neural network (Fig. 2).

The goal is to determine the optimal neural network structure for melanoma classification. In our previous work, AlexNet we used the same approach on clinical dataset [ 4 ]. In this work, a revised subset of images from ISIC, one of the Google InceptionV3 most widely used datasets for melanoma classification, were employed in an experimental study to test the pro- GoogleNet posed methodology. The convolutional neural network architecture uses a genetic algorithm which enables the population to evolve over successive generations in order Table 1 to obtain the best fitness. The initial generation of neural Performance drop after 100 training steps network (NN) is stochastic. As a result, the initial accuracy is remarkably low, but as the experiment progresses, the error attenuates due to a higher level of fitness in the NN population. In the most recent evolutionary iteration, In this context, we designed and deployed a hybrid a set of equivalent NN with high classification perfor- architecture based on Cloud, Fog, and Edge Computing mance was available. Our hybrid approach to melanoma to provide a Melanoma Detection service based on clindetection CNN design achieves 94% accuracy, 90% sensi- ical and dermoscopic images (see Fig. 3). Data buckets tivity, 97% specificity, and 98% precision. According to are kept in the cloud, and system training is executed. preliminary results, the proposed method could improve After each formation in the Fog area, where services are melanoma classification by eliminating the need for user executed, the orchestrator is in charge of distributing interaction and avoiding a priori network architecture the optimized services. Local calculations are performed selection. in the Edge area on IoMT devices (for example, smartphones). HiC-Otsu is a software component of the Fog system on the IoMT device that performs preliminary data analysis. To improve system performance, the QoS moderator annotates content. The generic user uses the output of services, but by loading data, he contributes to the system’s knowledge base. 3.1.4. Real Time melanoma detection support

Cardiovascular disease (CVD) is still the leading cause of

death worldwide. According to the World Health Organization (WHO), in 2017 more than 17.9 million people died from cardiovascular disease (31% of all deaths worldwide).

Premature Ventricular Contraction (PVC) is an additional heartbeat that occurs in one of two heart ventricles that delays the normal pumping order, first the atria, then the ventricles. PVC is of crucial importance in the cardiology ifeld, not only to improve the health system but also to reduce the workload of experts who analyze ECG manually. PVC is a non-harmful common occurrence represented highlight the latent patterns of non-PVC signal data that by extra heartbeats, whose diagnosis is not always easily indicate belonging to diferent arrhythmias. We first preidentifiable, especially when done by long-term manual processed the ECG signals with noise remove technique, ECG analysis. In some cases, it may lead to disastrous and then we created a matrix based on the wave distances consequences when associated with other pathologies. of each pair of analyzed images. This distance matrix as This work introduces an approach to classify PVCs using used for the next clustering analysis. One of the most machine learning techniques [ 7 ]. In particular, a group important goal of a clustering algorithm is to classify of six classifiers is used: Decision Tree, Random Forest, the data into a set of clusters in order to group similar Long-Short Term Memory (LSTM), Bidirectional LSTM, elements. K-means is a method based on the centroids, ResNet-18, MobileNetv2, and ShufleNet. Two types of the points belonging to the space of the features, that experiments are performed on data extracted from the mediates the distances between all the items belonging MIT-BIH Arrhythmia database: (i) the original dataset to the identified cluster. This algorithm chooses the k and (ii) the balanced dataset. MobileNetv2 came in first value in an arbitrary way, without knowing the classes in both experiments with high performance and promis- in the input dataset. The Elbow method was used to ing results for PVCs’ final diagnosis. The final results objectively choose the best value for k. As a result, we showed 99.90% of accuracy in the first experiment and were able to calculate the optimal number of clusters for 99.00% in the second one, despite no feature detection the explored data. In particular, this optimization on the techniques were used. The approach we used, which k-means analyzes the intrinsic pathological meaning of was focused on classification without feature extraction, non-PVC signals. Although these signals are not adeallowed us to dramatically reduce costs and computa- quately labeled in the dataset, we could identify possible tional times while providing excellent performance and common patterns connected to various arrhythmias in obtaining better results. Finally, this research defines as the non-PVC class. At present, our work is still in the a first step toward understanding the explanations for experimental phase. The identification of well-defined deep learning models’ incorrect classifications. clusters allows us to hypothesize the presence of morphological patterns common to diferent arrhythmias, in 3.2.2. Identification of a Pattern in Premature particular between PVC and non-PVC as shown in Fig. 4.

Ventricular Contractions Future studies will be able to help us better describe the clusters that exist in non-PVCs, making it easier to idenThe primary goal of this study is to find PVCs-related tify and label them based on their ECG signal feature patterns in ECG signals [ 8 ]. Specifically, our focus was to space.

The classification performance of the three neural networks was comparatively assessed on diferent datasets provided by two hospitals. This process allowed the evaluation of the robustness of the architecture versus the influence of diferent hardware instrumentation or acquisition protocols. The study concluded that the DCNN architecture performed better in terms of sensitivity and specificity and had the potential to reduce false positives.

Additionally, a technique called Grad-CAM was implemented to highlight pixels in all DBT slices of a given exam that were more relevant for the final classification task performed by the network. This technique could provide an indication of the position of the mass inside the slice(s) classified as abnormal and show the location of possible network activation in zones less relevant for the diagnostic task (see Fig.5).

3.3. Deep learning for breast cancer

detection 3.3.1. DCNN for cancer detection in DBT images Breast cancer is the most common type of cancer in women, and mammography is the most efective method for early detection. However, mammography has limitations in terms of sensitivity and specificity, especially for dense breasts. To overcome these limitations, new imaging technologies such as digital breast tomosynthesis (DBT) and breast computed tomography have been developed. DBT, in particular, ofers a (pseudo) threedimensional representation of the breast tissue and a clearer localization of possible lesions (masses and microcalcifications). However, interpreting a DBT exam requires the analysis of tens of image slices, adding complexity to the radiological clinical workflow. To improve experts’ performance in DBT exam analysis, computeraided detection (CAD) systems have been developed.

These systems help in managing the complexity of DBT lesion search space and potentially improve diagnostic accuracy. One specific goal of the CAD system is to achieve an acceptable trade-of between the computational costs arising from automatic analysis and classification performance. Previous studies have focused on developing CAD systems using hand-crafted features. In this study [ 9 ], a deep convolutional neural network for DBT images (DBT-DCNN) was developed to automatically classify the presence or absence of mass lesions in DBT exams.

The performance of this DCNN was compared to that of popular architectures (AlexNet and VGG19) (see Tab.2).

As discussed in Sec.3.3.1 DBT exam consists of tens of image slices, adding complexity to the radiological clinical workflow and, at the same time, representing a critical problem when approaching to DL solutions: building a dataset of medical images is a complex task due to privacy restrictions, the need for expert clinicians to report the images, and the costs and manual eforts required to process them. An additional challenge is the non-balancing of datasets, especially when a particular class is more Table 2 abundant than others. To address these challenges, new Evaluation in terms of classification absolute numbers, accu- techniques to augment and balance existing DBT datasets racy (acc) and sensitivity (s), of the DCNN architectures. with realistic synthetic samples become necessary. The TP TN FP FN acc s use of data augmentation techniques, and in particular (#) (#) (#) (#) (%) (%) of generative models, represents a possible solution to TL-AlexNet 940 249 208 9 84.6 99.0 the problem. In this work, we investigated an innovative TL-VGG19 832 215 242 17 74.5 87.7 solution in the field of generative models, in particular, DBT-DCNN 948 374 83 1 94.0 99.0 Generative Adversarial Network (GAN) models. However, GANs often sufer from training dificulties such as the problem of the gradient vanishing and mode collapse.

To cope with these problems, a new GAN architecture, known as Evolutionary GAN (E-GAN), has been designed that uses diferent metrics together to optimize the generator through an evolutionary approach. In this work [ 10 ] such a model has been applied to the problem of increasing the data of a DBT image dataset to generate a more significant number of samples of "sick" slices to balance the starting dataset. In the proposed E-GAN approach, a discriminator acts as an environment while a population of generators mutates to produce ofspring that adapt to the environment. Furthermore, once the current best discriminator is fixed, the quality and diversity of the samples generated are assessed, and only the best one for future training steps is kept. A first training session was performed on the entire dataset (results are shown in Fig.6) clearly showing some artifacts: non-continuity in borders, shadows, incoherence of the nipple etc. Although not definitive and usable for classification, the results represent a starting point for the development of future architectures in charge of 2.5D or even 3D. 4. Project • We proposed a PRIN on the thematic of reconstruction algorithms for 3D breast computed tomography; • Some of the works presented are part of the Artificial Intelligence in Medicine (AIM) project, funded by INFN.

References

[1] N. R. Abbasi , H. M. Shaw , D. S. Rigel , R. J. Friedman , W. H. McCarthy , I. Osman , A. W. Kopf , D. Polsky , Early diagnosis of cutaneous melanoma: revisiting the abcd criteria , Jama 292 ( 2004 ) 2771 - 2776 .

[2] L. Di Biasi , F. De Marco , A. A. Citarella , M. Castrillón-Santana , P. Barra , G. Tortora, Refactoring and performance analysis of the main cnn architectures: using false negative rate minimization to solve the clinical images melanoma detection problem , BMC Bioinformatics ( 2022 ).

[3] L. Di Biasi , F. De Marco , A. Auriemma Citarella , P. Barra, , S. Piotto Piotto , G. Tortora, Hybrid approach for the design of cnns using genetic algorithms for melanoma classification , in: 2nd International Workshop on Artificial Intelligence for Healthcare Applications (AIHA 2022 ), 2022 .

[4] L. Di Biasi , A. Auriemma Citarella , F. De Marco , M. Risi , G. Tortora, S. Piotto , Exploration of genetic algorithms and cnn for melanoma classification , in: J. J. Schneider , M. S. Weyland , D. Flumini , R. M. Füchslin (Eds.), Artificial Life and Evolutionary Computation , Springer Nature Switzerland, Cham, 2022 , pp. 135 - 138 .

[5] L. Di Biasi , A. A. Citarella , M. Risi , G. Tortora , A cloud approach for melanoma detection based on deep learning networks , IEEE Journal of Biomedical and Health Informatics 26 ( 2021 ) 962 - 972 .

[6] R. Francese , M. Frasca , M. Risi , G. Tortora , A mobile augmented reality application for supporting realtime skin lesion analysis based on deep learning , J. Real Time Image Process . 18 ( 2021 ) 1247 - 1259 .

[7] F. De Marco , F. Ferrucci , M. Risi , G. Tortora, Classification of qrs complexes to detect premature ventricular contraction using machine learning techniques , Plos one 17 ( 2022 ) e0268555 .

[8] F. De Marco , L. Di Biasi , A. A. Citarella , M. Tucci , G. Tortora, Identification of morphological patterns for the detection of premature ventricular contractions , in: 2022 26th International Conference Information Visualisation (IV) , IEEE, 2022 , pp. 393 - 398 .

[9] R. Ricciardi , G. Mettivier, M. Stafa , A. Sarno , G. Acampora, S. Minelli , A. Santoro , E. Antignani , A. Orientale , I. Pilotti , V. Santangelo , P. D'Andria , P. Russo , A deep learning classifier for digital breast tomosynthesis , Physica Medica 83 ( 2021 ) 184 - 193 .

[10] M. Stafa , L. D'Errico , R. Ricciardi , P. Barra , E. Antignani , S. Minelli , G. Mettivier, How to increase and balance current dbt datasets via an evolutionary gan: preliminary results , in: 2022 22nd IEEE International Symposium on Cluster, Cloud and Internet Computing (CCGrid) , 2022 , pp. 913 - 920 .