UniCas for Medicine and Healthcare

Abstract

With over twenty years of experience, our research group, afiliated with the Artificial Intelligence and Data Analysis Laboratory (AIDA), which belongs to the University of Cassino and Southern Lazio (UniCas), has been deeply engaged in artificial intelligence. The specific focus on Machine Learning, Pattern Recognition, and Deep Learning has evolved theoretically, with the development of specialized skills in model optimization, and practically, through application to real-world problems, particularly in the healthcare domain. In particular, attention is paid to designing and implementing Computer-Aided Diagnosis systems to support the prevention, diagnosis, and monitoring of Neurodegenerative diseases, Specific Learning Disorders, breast cancer, diabetic retinopathy, and movement-related disorders. Diferent data is utilized to reach the objectives of the AIDA Lab, and many approaches are implemented. Handwriting analysis is exploited to support the diagnosis of neurodegenerative diseases and specific learning disorders and to monitor them over time. Handwriting analysis encompasses two distinct approaches: examining dynamic features and scrutinizing handwriting sample images. This comprehensive approach allows a more thorough understanding of the individual's writing characteristics. Additionally, in Neurodegenerative Diseases, advancements include the utilization of 3D image analysis of MRI scans to aid in the detection of Alzheimer's disease, further enhancing diagnostic capabilities in this field. Mammograms are used for breast cancer prevention and diagnosis, while retinal images are used for diabetic retinopathy detection, particularly focusing on detecting small lesions. Detecting small lesions is a crucial step in diagnosis, as is identifying microcalcifications in digital mammograms and microaneurysms in digital fundus images. To address this challenge, we propose a novel architecture called GravityNet. Another field of study in the research conducted by the AIDA Lab is movement analysis, focusing on gait analysis, enabling precise evaluations in real-world settings and potential applications in Parkinson's disease assessment. To this end, Machine and advanced Deep Learning techniques are employed, such as deep cascades of boosting classifiers and Deep Convolutional and Attentional Neural Networks.

Keywords

• eol>Neurodegenerative Diseases
• Handwriting
• 3D Image Analysis
• Breast Cancer
• Small Lesion
• Movement Analysis
• Retinopathy

1. Neurodegenerative Diseases and Brain Disorders

The prevalence of neurodegenerative diseases (NDs) has been steadily increasing in recent years, underscoring a concerning trend. Another aspect underscoring the importance of research in this field is that NDs currently lack a cure. They can cause cognitive impairments manifesting as dificulties in memory, language, thinking, judgment, and motor skills. Individuals exhibiting a combination of these symptoms face a significantly heightened risk of developing dementia and, in more severe cases, Alzheimer’s Disease (AD) or Parkinson’s Disease (PD). Given the progressive nature of these conditions, early detection becomes crucial to initiate therapies to mitigate their efects.

Concerning other brain-related disorders, the prevalence of specific learning disorders (SLD) presents a significant and persistent challenge in educational and developmental psychology. These disorders encompass a range of conditions afecting an individual’s ability to acquire and air movements up to a maximum height of 3 from apply foundational academic skills, such as reading, writ- its surface. After the acquisition step, many images of ing, and mathematics, despite adequate intelligence and the handwritten traces were collected or generated. Paopportunities for learning. Unlike NDs, which primar- per sheets were scanned to generate the ofline image ily afect cognitive functioning in adulthood, specific dataset [ 2 ], while the raw data, expressed as coordinates, learning disorders often manifest during childhood and were used to generate synthetic images by interpolating can persist into adolescence and adulthood if left un- consecutive points. In this way, we generated the binary treated. Given the chronic nature of specific learning dataset [ 3 ] and the RGB one [ 4 ], which encoded dynamic disorders and their potential to impact multiple domains information in the three color channels. Figure 1 illusof functioning, early identification and intervention are trates the various images examined and produced within paramount. this research path. In particular, they refer to the Clock Drawing Test.

1.1. Handwriting

Handwriting analysis has emerged as a valuable tool in the early detection of NDs like AD. This method’s significance lies in its ability to provide insights into various cognitive, memory, and motor functions integral to writing. The complexity of writing, which encompasses language proficiency, executive skills, and attention span, can reveal changes in linguistic complexity, word choice, and syntactic structures. Such alterations are pivotal for early AD diagnosis as they reflect the underlying cognitive changes. Additionally, the link between writing and memory ofers a window into the patient’s ability to associate names with concepts, identify temporal relationships, and recall previously seen words, highlighting declines in episodic memory. Moreover, the analysis of ifne motor control through handwriting, such as changes in pressure, speed, and stroke patterns, can indicate early declines in motor functions. These aspects of handwriting analysis serve as critical markers for the early identification and intervention of AD, aiming to delay its progression and improve patient outcomes.

These datasets were used to feed a system that exploited the ability of Convolutional Neural Networks (CNNs) to extract features automatically. Various experiments were conducted using diferent task configurations and datasets. Despite deploying numerous experimental settings, they all shared a common baseline architecture, facilitating comparative analysis and yielding insightful results. The baseline experimental architecture comprises four key steps: data acquisition, feature extraction/engineering, Machine Learning (ML) and Deep Learning (DL) classification, and combining rule application. The outcomes of this research revealed interesting trends. Across various tasks, CNN-extracted deep features consistently outperformed other approaches, em1.1.1. Task Images Analysis phasizing the advantage of DL. Experimentation with In the context of AD, alterations in handwriting are of- ofline images suggests their potential in diagnosing NDs, ten observable due to the cognitive and motor changes underscoring the importance of considering shape details associated with the condition. In the 2000s, researchers in distinguishing patients from healthy individuals, espebegan to recognize the potential of handwriting analysis cially in medical applications where sensitivity is crucial. as a non-invasive and accessible tool for early detection. Finally, this research highlighted that some handwriting In 2018, researchers from the AIDA LAb started a metic- tasks were more significant in supporting AD diagnosis, ulous data acquisition campaign by administering an and combining them improved the final prediction. experimental protocol [ 1 ] of 25 handwriting tasks. The research included a total of 174 participants, with 89 1.1.2. Dynamic Analysis patients diagnosed with AD and 85 healthy control subjects. Each participant was invited to perform the exper- Our research also focused on handwriting analysis by imental protocol by using the WACOM Bamboo Folio capturing raw data through digital devices, which track graphic tablet, enabling participants to write on standard intricate handwriting details like pressure, tilt, and moA4 white paper sheets using a pen that appears typical. tion. This data was meticulously organized and analyzed This pen not only produced ink traces on the sheet but to extract critical dynamic and static features, such as also generated digital information recorded by the tablet velocity, acceleration, and stroke pressure, providing a in the form of spatial coordinates and pressure for each rich dataset for examination. Before the ML algorithms, point (, , and ). The tablet additionally captured in- we conducted thorough preprocessing to normalize and clean the data, followed by feature selection processes [ 5 ] to highlight the most indicative characteristics of the handwriting. This preparation phase also included hyperparameter tuning to refine our ML models for optimal performance. Our methodology incorporated two distinct ML approaches: a Statistical Approach, analyzing overall handwriting patterns to detect deviations indicative of diseases, and a Stroke Approach, focusing on the detailed analysis of individual handwriting strokes. This dual perspective enabled us to capture both the broad characteristics and the fine-grained patterns in handwriting that may signal the early stages of conditions like Alzheimer’s disease. By leveraging digital technology for data acquisition and applying a nuanced ML analysis, our study aims to enhance the early diagnosis and understanding of NDs, ofering the potential for timely intervention and improved patient outcomes.

1.2. Specific Learning Disorder

Original 3D MRI

Explainable 3D MRI

Handwriting and drawing are essential for preschool and 1.3. 3D Images Analysis school-age children’s cognitive and motor development.

Studies indicate that a significant portion of the school 3D imaging is a widely used technique in diagnostic proday, ranging from 30-60%, is dedicated to fine motor activ- cedures for brain disorders. Magnetic resonance imaging ities, primarily handwriting. Graphomotor tasks, crucial (MRI) provides a non-invasive means to observe and anfor academic success and personal expression, account alyze in vivo pathological changes in the brain related for about 40% of a primary school student’s activities. to Brain Disorders, facilitating the study of disease evoIn this age group, 20% of children are at risk of expe- lution. MRI analysis is significant in AD diagnosis since riencing graphomotor dificulties, with 27% displaying it is identified by structural and functional changes in subpar graphomotor abilities. It is essential to note the dynamically changing morphological patterns, which role of Specific Learning Disorders (SLDs), such as dys- are appropriately captured with high-resolution MRI. It graphia, which are closely associated with graphomotor is also noteworthy that brain atrophy, a distinctive AD challenges. Early detection and intervention are crucial symptom, can be identified through MRI. This form of in addressing these disorders. Our research aims to de- atrophy serves as a reliable marker of the disease. It is velop innovative ML solutions that promptly identify and indicative of its progression, as well as being associated address handwriting issues early in a child’s academic with tau deposition and neuropsychological impairments, journey. Timely intervention is crucial, as the window essential factors in the clinical manifestation of AD. for efective rehabilitation narrows over time. ML is es- In recent years, CNNs and Vision Transformers have sential for identifying subtle yet vital changes in the transformed neuroimaging data analysis for AD, moving graphomotor cycle that experienced specialists may miss. beyond the traditional ML that focuses on approaches The study started with creating an experimental protocol that rely on handcrafted features and classifiers. Recent to reveal the fundamental features of handwriting. The works have proven promising in diagnosing and predicthandwriting was digitized using graphic tablets, generat- ing Alzheimer’s; however, at the same time, presenting ing a comprehensive dataset for training ML models to diferent limitations mainly linked to the use of 3D CNN accurately diferentiate between various learning impair- networks. Training these networks requires many samments. The investigation will be longitudinal, spanning ples, often unavailable in the medical field. One way four years of data collection. This will increase the depth to mitigate this problem is to use 3D volume slicing apof the dataset and allow for a detailed exploration of proaches. The volume is divided into 2D slices by slicing the evolution of graphomotor skills and associated chal- from one of the three anatomical body planes. This solulenges. The monitoring will enable the study to identify tion partially solves the problem but involves the loss of graphomotor problem trends and indicators, enhancing learning global patterns of the entire volume. ML algorithms’ capacity to classify handwriting issues Another key aspect of DL models in the medical field and lead to accurate early interventions. is the interpretability of their results. It is crucial for clinical applications and remains a significant challenge that hinders the deployment of DL-aided diagnostic systems in real-world scenarios.

Our research aims to develop models that compensate for single calcification detection in mammograms involvfor current limitations. We are developing approaches ing a specialized convolution layer operating in the frethat allow us to be efective with few training samples quency domain as the first layer of a CNN [ 7 ]. This layer and are explainable at the same time. Radiologists typ- automatically learns a bank of Diference of Gaussians ically examine 3D images as a series of 2D images and (DoG) band-pass filters parameterized by their associated base their diagnosis by focusing on specific brain parts pairs of standard deviations inversely proportional to the from diferent views. Inspired by this, we have developed iflter’s bandwidth. Additionally, we explore the potential a diagnosis and XAI approach that allows us to diagnose of transformer models in backbone-head architecture for AD and, at the same time, highlight the areas predomi- lesion detection [ 8 ]. nantly afected by the disease. From a series of 2D slices Amalgamating traditional image processing techextracted from the volume, the network can reconstruct niques with ML and DL methodologies has significantly a three-dimensional saliency map highlighting the areas advanced cancer detection in breast imaging. Ongoing afected by AD with voxel-level precision (see Figure 2). exploration of innovative approaches, such as frequencyPreliminary results show efectiveness in highlighting iflter-integrated CNNs and transformer-based models, areas commonly associated with the disease, suggesting ofers promise for further enhancing detection accuracy. that these approaches are promising for diagnosing and predicting brain disorders.

2.2. Small lesion detection

2. Medical Image Analysis

2.1. Breast cancer detection

Breast cancer ranks as the most prevalent form of cancer among women and stands as the second leading cause of death. Detecting breast cancer at an early stage is crucial for increasing survival rates, prompting the implementation of mammography screening programs in numerous countries. Mammograms can reveal various abnormalities, including calcifications, masses, architectural distortions, and focal asymmetry.

Computer-aided diagnosis (CAD) systems have been devised to aid radiologists in analyzing mammographic images for diagnostic and screening purposes. Incorporating ML and DL techniques has led to significant progress in numerous tasks, such as image classification, lesion detection, and segmentation. CNNs stand as the de facto standard in many image analysis applications, demonstrating the ability to detect microcalcifications and other types of lesions accurately. Our research revolves around utilizing CNNs and transformer models to address various challenges in mammogram analysis.

Transformer-based models, such as the Vision Transformer, promise to capture a broader context in breast imaging tasks. Ongoing research on transformer-based models in mammography indicates their eficacy in tasks like multi-view mammogram classification.

One of our research interests involves the classification of entire mammograms as malignant or normal, assessing their performance based on the type of lesion present and at diferent input resolutions [ 6 ]. Mammographic images exhibit lesions with diverse characteristics in size, shape, texture, and sparsity, which may align better with local convolutional paradigms or self-attention capable of capturing long-range features. We also propose a method There has been significant interest in utilizing DL techniques for medical image analysis in recent years. These methods, employing advanced neural network architectures, have greatly improved the ability to extract valuable information from complex medical images. This advancement holds promise for more accurate diagnostics and tailored healthcare solutions. While current object detection models have shown impressive success, the increasing resolution of medical images presents a unique challenge, requiring greater attention to detail and identifying smaller objects within the images. Small lesions can be critical indicators of various medical conditions, making their accurate detection crucial for diagnosis and treatment planning.

To address this, a novel DL model, called GravityNet [ 9 ], is proposed to enhance the detection of small objects, employing innovative anchoring techniques known as gravity points, that appear to be ’attracted’ to lesions.

The architecture comprises a backbone network and two subnets dedicated to regression and classification tasks.

This proposed approach notably enhances detection accuracy, leading to earlier diagnoses and improved patient outcomes. The results underscore the significance of this innovative approach, demonstrating its broad implications across various clinical applications.

For instance, GravityNet is applied to identify calcifications in digital mammograms, which are crucial for the early detection of breast cancer. Furthermore, its efectiveness in detecting microaneurysms in retinal images could significantly aid the early diagnosis and management of diabetic retinopathy, a leading cause of blindness worldwide.

In addition, GravityNet shows promise in cytology imaging, particularly in the analysis of whole slide images to detect cell nuclei, which is crucial for cervical cancer identification. Accurate identification of cell nuclei in cytology images helps pathologists identify morphology and abnormal cell patterns indicative of cervical cancer.

This capability has significant potential to improve the eficiency and accuracy of cervical cancer screening programs, enabling earlier diagnosis and intervention to reduce the morbidity and mortality associated with this prevalent cancer.

Finally, GravityNet extends to the identification of large vessel occlusions (LVOs), a crucial task in stroke diagnosis and treatment planning. LVOs are obstructions in the brain’s main arteries, and their early detection is essential to determine appropriate treatment strategies to minimize neurological damage and improve patient outcomes. Leveraging its advanced DL algorithms, GravityNet can accurately detect and localize LVOs in volumetric space within computed tomographic angiography (CTA) scans. This capability allows healthcare professionals to quickly identify patients at risk of a severe stroke and promptly initiate interventions such as thrombectomy or thrombolysis, which can significantly reduce disability and mortality rates. need for markers attached to the patient’s skin, streamlining the experimental workflow and reducing preparation time. Moreover, these systems allow to perform GAn even outside gait laboratories. This makes them useful for biomechanical evaluations in sports settings and assessing ambulatory functions in patients with neuromotor impairments during routine activities. This approach combines movement and GAn with clinical precision and practical applicability, which allows for patient care and athletic performance enhancement in real-world settings.

MMA systems bridge the gap between rigorous laboratory analysis and everyday assessment, representing a 3. Movement Analysis significant stride in precision medicine and rehabilitative care. Figure 3 shows the practical application of an innoGait Analysis (GAn) is an objective assessment of a per- vative MMA framework developed to estimate joint and son’s walking abilities and is an essential part of mo- segment angles from 2D videos and images. It showcases tor assessment. It helps health professionals make in- real-time data acquisition and analysis during a treadformed clinical decisions and develop targeted rehabil- mill gait assessment, proving the successful integration itation strategies to improve gait functions. In clinical of technology and biomechanics in modern therapeutic practice, GAn is performed by healthcare professionals and athletic environments. By delivering precise and via standardized questionnaires, functional tests, and vi- dependable joint and segment dynamics measurements, sual observation of the patient’s walking pattern. The such a solution can significantly augment human moveidentification of gait irregularities is based on the sub- ment analysis, potentially elevating health outcomes and jective quantification of spatio-temporal parameters and athletic performance. detailed kinematic and kinetic evaluations. Moreover, in the context of Parkinson’s Disease (PD),

Various advanced technologies have completely a common neurodegenerative disorder with progressive changed the way of analyzing motion objectively. Op- loss of dopaminergic and other sub-cortical neurons, GAn tical motion capture systems are the most accurate and is an essential tool for the diagnosis and progression of precise for joint kinematics assessment, making them the the disease, which is characterized by motor symptoms "gold standard" within laboratory settings. Such tech- such as tremors, rigidity, bradykinesia (slowness of movenologies enable the accurate acquisition of motion data ment), and postural instability. Through the utilization using reflective markers placed at strategic joint locations of GAn, clinicians can better understand how PD afects and multiple cameras that track them. However, they an individual’s mobility in daily life, enabling the identiare mainly restricted to specialized gait laboratories and ifcation of specific gait patterns associated with PD proresearch settings due to their limitations, which include gression. To achieve this aim, ML algorithms are used to high costs, technical expertise demands, and extensive classify diseases based on spatiotemporal and kinematic setup durations, hampering widespread clinical adoption. gait features extracted from MMA. An essential aspect ML approaches are increasingly used in GAn, developing of employing ML in medical diagnostics is explainability, from integrating computational intelligence with biome- particularly when understanding the impact of diferent chanics to ofer robust methods for analyzing complex features on the model’s predictions. Understanding what movement patterns [ 10 ]. CNN are increasingly utilized features have a major impact on the outcome is essential for human pose estimation within marker-less motion in the medical field. The use of MMA systems presents a analysis (MMA) systems, significantly augmenting their significant advancement in the management of PD, enprecision and reliability. In contrast to marker-based abling continuous and real-time monitoring of patients’ motion analysis systems, MMA systems eliminate the movements. This continual observation facilitates the development of highly personalized rehabilitation and therapy programs that can be adjusted daily to meet the evolving needs of each patient.

Acknowledgments

References

[1] N. Cilia , C. De Stefano , F. Fontanella , A. Scotto di Freca, An experimental protocol to support cognitive impairment diagnosis by using handwriting analysis , Procedia Computer Science 141 ( 2018 ) 466 - 471 .

[2] N. Dalia Cilia , T. D'Alessandro , C. De Stefano , F. Fontanella , Ofline handwriting image analysis to predict alzheimer's disease via deep learning , in: 2022 26th International Conference on Pattern Recognition (ICPR) , 2022 , pp. 2807 - 2813 . doi: 10 .1109/ICPR56361. 2022 . 9956359 .

[3] N. D. Cilia , T. D'Alessandro , C. De Stefano , F. Fontanella , M. Molinara , From online handwriting to synthetic images for alzheimer's disease detection using a deep transfer learning approach , IEEE Journal of Biomedical and Health Informatics 25 ( 2021 ) 4243 - 4254 . doi: 10 .1109/JBHI. 2021 . 3101982 .

[4] N. D. Cilia , T. D'Alessandro , C. De Stefano , F. Fontanella , Deep transfer learning algorithms applied to synthetic drawing images as a tool for supporting alzheimer's disease prediction , Machine Vision and Applications 33 ( 2022 ) 49 .

[5] T. D'Alessandro , C. De Stefano , F. Fontanella , E. Nardone , A. Scotto di Freca, Feature evaluation in handwriting analysis for alzheimer's disease using bayesian network , in: International Graphonomics Conference, Springer, 2023 , pp. 122 - 135 .

[6] M. Cantone , C. Marrocco , F. Tortorella , A. Bria , Convolutional networks and transformers for mammography classification: an experimental study , Sensors 23 ( 2023 ) 1229 .

[7] M. Cantone , C. Marrocco , F. Tortorella , A. Bria , Learnable dog convolutional filters for microcalcification detection , Artificial Intelligence in Medicine 143 ( 2023 ) 102629 .

[8] A. S. Betancourt Tarifa , C. Marrocco , M. Molinara , F. Tortorella , A. Bria , Transformer-based mass detection in digital mammograms , Journal of Ambient Intelligence and Humanized Computing 14 ( 2023 ) 2723 - 2737 .

[9] C. Russo , A. Bria , C. Marrocco , Gravitynet for end-to-end small lesion detection , Artificial Intelligence in Medicine ( 2024 ) 102842 . doi: 10 .1016/j. artmed. 2024 . 102842 .

[10] S. Galasso , R. Baptista , M. Molinara , S. Pizzocaro , R. S. Calabrò , A. M. De Nunzio , Predicting physical activity levels from kinematic gait data using machine learning techniques , Engineering Applications of Artificial Intelligence 123 ( 2023 ) 106487 . URL: https://www.sciencedirect.com/science/ article/pii/S0952197623006711. doi:https://doi. org/10.1016/j.engappai. 2023 . 106487 .