Pilots

Experiments in Montenegrin S3 Domains

Several experiments were implemented towards piloting the use of HPC and AI in priority domains of S3 Montenegro. We involved young researchers at different levels (PhD, MSc, and BSc) into experimenting through direct engagement in activities on the project or through mentoring and training processes implemented through HPC4S3ME and/or project cross-collaboration. Below, you can find descriptions of some of the pilot activities and results that ended up being published either in the form of MSc/BSc theses or publications and presentations at international conferences.

Optimization of Solar Energy Management through AI-based Systems

S3 Domain: Energy

Summary: In the contemporary pursuit of sustainable energy solutions, solar energy has emerged as a prominent candidate. Harnessing solar power through solar panels not only reduces dependency on fossil fuels and lowers carbon dioxide emissions but also offers multiple economic benefits. However, optimizing the use of solar panels requires precise management and forecasting of solar energy production. This project explores the optimization of solar energy usage and the management of solar panels in hybrid energy systems using AI models. A model has been developed to predict solar radiation, enabling users to plan their energy needs and maximize savings. The project encompasses the collection of weather data, energy usage optimization, and proposals for practical implementation.

Challenges Addressed: Precise management and forecasting of solar energy production are critical in hybrid energy systems. Users often struggle to decide when to use solar energy versus grid energy, especially during periods of low solar radiation. The lack of an integrated system that can predict monthly solar radiation makes it difficult to optimize the use of solar panels and can lead to energy deficits or losses. Additionally, users need real-time recommendations for efficient energy usage, which requires complex models for analyzing and predicting weather patterns. These are exactly the challenges that the proposed model aims to solve.

Proposed Solution: The proposed solution involves the development of an application that uses Long Short-Term Memory (LSTM) neural networks to predict solar radiation. The system collects weather data and uses it to generate accurate predictions of solar energy production. The recommended system includes both web and mobile applications, allowing users to monitor data in real time and receive recommendations on when to use solar energy or switch to the grid. The system also supports integration with IoT devices for improved functionality and accuracy. Recommendations will be based on analyses using a combination of CNN and LSTM models to analyze spatiotemporal patterns in the data.

Business Impact: The implementation of the proposed system can significantly enhance energy management for both end-users and companies that sell solar equipment. Accurate predictions enable users to optimize their use of solar energy, reducing dependence on the grid and lowering energy costs. Companies can use these predictions to better plan sales and implement solar solutions for clients. The system also supports agricultural users in planning their energy needs for irrigation and machinery. On a global scale, improved solar energy utilization contributes to reducing CO2 emissions and promoting sustainable energy practices.

Potential Benefits:

• Increased Energy Efficiency: Optimizing the use of solar energy reduces dependence on fossil fuels and lowers energy costs.
• Economic Savings: Users can reduce electricity bills and generate income by selling excess energy.
• Support for Sustainable Practices: Promotes sustainable energy practices and reduces CO2 emissions.
• Real-Time User Recommendations: Enables users to make informed decisions about energy management.
• Integration with IoT: Enhances system functionality and accuracy through connectivity with smart devices.

Embracing AI for Smart Hospitality

S3 Domain: Tourism

Summary: The tourism industry, particularly in Montenegro, is undergoing a digital transformation fueled by artificial intelligence, blockchain, and high-performance computing. A standout innovation is the integration of chatbots in hotels to improve guest services and streamline reception processes. This study showcases a chatbot developed for a Montenegrin hotel using the Voiceflow platform, aimed at enhancing the guest experience by reducing wait times and offering real-time assistance with a dedicated knowledge base for each hotel.

Challenge Addresed: Hotels face increasing demands for seamless, efficient, and personalized guest interactions. Reception staff often deal with repetitive queries, leading to delays in services and limited availability for other guest-centric tasks. To maintain high-quality service standards, there is a need for innovative solutions that can manage common inquiries without diminishing the personal touch guests expect.

Proposed Solution: The study proposes implementing a hotel chatbot designed to handle frequent guest queries, empowering receptionists to focus on more complex or personalized requests. This chatbot, built on the Voiceflow platform, utilizes a comprehensive knowledge base tailored to each hotel’s unique policies and frequently asked questions. It is integrated with web and messaging platforms, providing a user-friendly interface for guests to interact with, whether it’s checking room availability or exploring on-site activities.

Business Impact: The adoption of a hotel chatbot system drives substantial business impact by optimizing operations and boosting customer engagement. By automating routine guest inquiries, staff are freed to address high-priority interactions, improving service quality and resource management. This scalability reduces the need for additional seasonal staff, cutting labor costs while ensuring prompt service during peak times. Additionally, the chatbot’s data analytics provide valuable insights into guest preferences, enabling hotels to tailor offerings and enhance customer satisfaction. This technology positions hotels at the forefront of digital transformation in tourism, delivering a competitive advantage and reinforcing a modern, tech-savvy brand image.

Potential Benefits

• Enhanced Efficiency: Reduces reception staff workload by automating common queries.
• Improved Guest Experience: Delivers accurate information quickly, enhancing guest satisfaction.
• Customization: Each hotel can tailor the chatbot to align with its specific services and regulations.
• Scalability: Capable of handling numerous inquiries simultaneously, supporting high-traffic periods effortlessly.
• Data Security: The platform incorporates robust security protocols to protect guest information.

Preserving Cultural Heritage with AI

S3 Domain: Tourism

Summary: This pilot explores how artificial intelligence (AI) and computer vision can be used to preserve and promote Montenegro’s rich cultural heritage. It focuses on developing a system that can recognize unmarked monuments and provide information about them, enhancing the tourist experience and contributing to the preservation of often overlooked historical sites.

Challenge Addressed: Montenegro has a wealth of historical sites, many of which are unmarked or poorly maintained. This makes it difficult for tourists and even locals to appreciate the historical significance of these sites. The project addresses this challenge by using AI to identify and provide information about these hidden historical gems.

Proposed Solution: The project uses a combination of YOLOv8, a state-of-the-art object detection model, and Flask, a web framework, to create an application that can recognize monuments from images. The application allows users to upload an image of a monument and receive an annotated image along with a description. This information helps users understand the historical significance of the monument.

Potential Benefits

• Enhances the tourist experience by providing historical context to unmarked monuments.
• Helps preserve Montenegro’s cultural heritage by raising awareness of forgotten sites.
• Positions Montenegro as a leader in the use of AI for tourism and cultural heritage preservation.
• Contributes to the achievement of the government’s tourism development goals.

Cattle Detection using Edge Devices

S3 Domain: Agri-Food

Summary: This polot focuses on the development of a system for real-time detection and tracking of livestock using cameras connected to a Jetson Nano device and the YOLO v8 model for animal recognition. Implementing this system enables precise tracking of the movement and condition of livestock, facilitating farm management and supervision. The system provides real-time notifications and key information for efficient resource management.

Challenges addressed: Agriculture is a key sector for the economy of many countries, including Montenegro. However, this sector faces numerous challenges that can negatively impact productivity and efficiency. High labor costs are a significant problem as farmers often have to hire a lot of workforce for livestock monitoring and management. Traditional methods of livestock tracking, such as manual counting and visual supervision, are often costly, time-consuming, and prone to errors. Inefficiencies in resource management can lead to losses and reduced profitability. For instance, if livestock health problems are not detected in time, it can result in higher treatment costs and reduced production. The lack of modern technologies further exacerbates the situation as many farmers do not have access to tools that could significantly improve their operations and ease their daily lives in rural areas. The implementation of advanced (smart) technologies such as edge devices and machine learning algorithms can significantly enhance efficiency and productivity in agriculture. The livestock detection and tracking system uses cameras connected to the Jetson Nano device and the YOLO v8 model for animal recognition.

Proposed solution: The proposal is to develop a real-time livestock detection system using cameras connected to a Jetson Nano device and the YOLO v8 model for animal recognition. The entire training process can be carried out using HPC servers and tools, which would significantly speed up the model development process. The goal is to ease farmers’ supervision and management of livestock, reduce labor costs, and increase farm safety. The system allows real-time tracking and provides accurate information on the movement and condition of livestock, which is crucial for timely decision-making and optimal resource management. The Jetson Nano device is a compact and energy-efficient computer that enables local data processing, reducing the need for transmitting large amounts of data to remote servers. Cameras connected to the Jetson Nano device record video in real time while the YOLO v8 model processes the footage and recognizes livestock. The livestock detection system enables precise tracking of livestock movement and condition. For example, if a cow strays from the herd or stays in one place longer than usual, the system can send a warning to the farmer. This allows for quick reactions and prevents potential problems before they become serious. Implementing this system can lead to significant savings and increased productivity, directly improving the economic sustainability of agricultural enterprises.

Business impact: The development and implementation of a livestock detection system can have a significant positive impact on the agricultural sector. Early livestock detection enables less invasive and costly treatments, resulting in overall savings for the agricultural sector. For example, if a health problem is detected early, treatment costs will be lower and recovery faster. Fast and efficient diagnosis through an AI system allows farmers to optimize their resources and increase productivity. The system provides real-time information crucial for timely decision-making.

Potential Benefits:

• Improved efficiency and safety in agricultural enterprises.
• Reduced labor costs and increased productivity. Automating livestock supervision reduces the need for an additional workforce, leading to significant savings.
• Providing real-time information for optimal resource management. The system allows farmers to quickly respond to changes and make informed decisions, improving resource management.
• Enabling earlier detection of livestock problems. Early detection of diseases or unusual behavior can lead to correct and timely reactions.

Forecasting Icterus Using Machine Learning

S3 Domain: Medicine

Summary: This pilot explores the application of machine learning for the classification of types of icterus (jaundice). Various algorithms were examined to achieve the most accurate diagnosis, which is crucial for appropriate patient treatment. After data preprocessing, the models were evaluated using standard metrics such as accuracy, precision, recall, and F1 score. The Logistic Regression model achieved an accuracy of 95.33%, with a precision of 95.85%, a recall of 95.33%, and an F1 score of 95.31%. The Naïve Bayes model performed exceptionally well across all metrics, achieving 99.5%, though it showed potential signs of overfitting. The Support Vector Machine models had varying results, with the Linear SVM achieving an accuracy of 94.5%, the Radial SVM at 79.17% and the Polynomial SVM at 79.33%. The Multi-layer Perceptron Classifier, particularly the model with five hidden layers, delivered the best performance, achieving an accuracy of 96%, precision of 96.33%, recall of 96.0%, and an F1 score of 95.99%. While the Naïve Bayes model demonstrated high accuracy, further work is necessary to mitigate the risk of overfitting.

Challenges addressed: Icterus is a medical condition characterized by yellowing of the skin and sclera due to the accumulation of bilirubin in the body. There are three main types of icterus: extrahepatic, intrahepatic, and prehepatic. Accurate diagnosis of the type of icterus is crucial for proper treatment. However, precise classification can be challenging due to the similarity of symptoms among the types and the need for quick decision-making. Conventional diagnostic methods often rely on subjective assessments and limited laboratory tests, which can lead to incorrect diagnoses and inadequate treatment.

Proposed solution: This paper explores various machine learning algorithms for classifying types of icterus using a customized dataset. The dataset used in the research comes from the Verlab Institute and is not publicly available. It contains data on patients with icterus (jaundice) and is divided into three main categories based on the type of icterus: extrahepatic, intrahepatic, and prehepatic. The dataset consists of 3000 instances with 11 attributes.

Business impact: Implementing such a system in medical institutions can significantly improve the process of diagnosing icterus. Automating the diagnostic process can reduce the need for lengthy and expensive laboratory tests, enabling faster and more accurate patient treatment. Precise classification of icterus allows doctors to immediately take appropriate therapeutic measures, reducing the risk of complications and improving treatment outcomes.

Potential Benefits

• Faster diagnosis: automation reduces the time needed for diagnosis.
• Higher accuracy: machine learning reduces subjective errors in diagnosis.
• Cost reduction: less need for expensive laboratory tests.
• Improved treatment outcomes: more accurate diagnosis enables more appropriate treatment.
• Advancement in medicalrResearch: collected data can help in further research and the development of new therapies.

Ovarian Cancer Detection

S3 Domain: Medicine

Summary: In the fight against ovarian cancer, science is making significant strides. With the help of artificial intelligence (AI) and computer vision technology to develop a system that automatically detects tumors in medical images, this system could improve early diagnosis and treatment precision. The project involves testing various AI models, such as YOLOv8 and YOLOv7, to identify the one with the best performance, aiming to enhance the accuracy of ovarian cancer detection and enable better outcomes for patients.

Challenges Addressed: Ovarian cancer poses a significant challenge to women’s health, and current diagnostic methods are limited in their accuracy and speed in detecting the disease. Transvaginal ultrasound, a commonly used method, can miss small tumors or mistakenly identify them as benign cysts. Additionally, blood tumor markers, such as CA-125, can be elevated in women with benign conditions and do not provide definitive evidence of cancer. These shortcomings often lead to delays in diagnosis, meaning that cancer is detected in later stages when treatment is more difficult and less successful. Moreover, traditional diagnostic methods can be subjective and prone to human error. Different interpretations of results by doctors can lead to misdiagnosis or missed diagnoses. Therefore, current screening requires significant time and expertise, which limits its availability and effectiveness, especially in rural areas.

Proposed Solution: The development of an innovative AI system for detecting ovarian cancer based on advanced machine learning technologies is proposed. This system would utilize YOLOv8 and YOLOv7 models to analyze medical images of the ovaries. The model would be trained on large datasets of labeled medical images, enabling it to develop the ability to automatically recognize tumors.

Business Impact: The development and implementation of an AI system for detecting ovarian cancer could greatly benefit the healthcare sector by reducing costs, boosting productivity, improving patient care, and creating new market opportunities. By enabling early detection, the system allows for less invasive and more affordable treatments, like early-stage surgery and radiation, which translates into substantial savings for the healthcare system. Additionally, rapid and accurate diagnosis empowers doctors to see more patients and spend more time with each, thereby enhancing productivity and efficiency in healthcare facilities. For patients, early-stage cancer detection leads to better treatment outcomes, higher chances of recovery, and improved satisfaction, which, in turn, enhances the reputations of healthcare institutions. Finally, this AI-driven innovation opens new opportunities for companies in both the medical and technology fields, fostering the development of market products and services such as AI model maintenance, medical image analysis software, and cancer screening services.

Potential benefits

• Improved diagnostics and potential for early detection
• Improvement in Patient Quality and Safety: Patients would have greater confidence in their doctors, knowing they are receiving accurate and timely diagnoses.
• Increased Efficiency of Healthcare Facilities: Immediate image analysis would accelerate the process of detecting potential ovarian tumors.
• Cost Reduction for Patients: Early detection of ovarian cancer offers the possibility of successful treatment, reducing the costs of prolonged treatments for patients

Detection of Scoliosis

S3 Domain: Medicine

Summary: This research presents the development of a model for scoliosis detection using artificial intelligence (AI) and computer vision alongside HPC (High-Performance Computing). The application of the YOLOv8 algorithm enables precise recognition of scoliosis in spinal X-ray images, improving diagnostics and allowing for quicker patient treatment response, while HPC enables fast training of the model itself and provides excellent resources. A web interface has also been developed, allowing users easy access to image analysis and obtaining recommendations for further treatment.

Challenges addressed: Scoliosis is a spinal deformity that can lead to serious health problems if not detected and treated in time. Traditional diagnostic methods are often subjective and depend on the expertise of the doctor, which can lead to delays in detection and treatment. There is a need to develop a faster, more precise, and objective method for scoliosis detection that could be applicable in rural areas with limited access to specialized medical equipment and experts.

Proposed solution: An innovative AI system based on deep learning for scoliosis detection has been developed. Using a dataset of 245 images from the Roboflow platform, the model has been trained for the automatic recognition of spinal deformities. The web application allows users to analyze scoliosis detection results through a simple interface. Tools such as HTML, CSS, JavaScript, and Flask were used, while MySQL database was used for data management.

Business impact: The implementation of this AI system can significantly reduce healthcare costs by enabling earlier diagnosis and less invasive treatments. Increased productivity of healthcare workers, better patient care, and opening new market opportunities for medical and technological companies are key advantages of this solution.

Project Benefits:

• Improvement of patient quality and safety.
• Increased efficiency of healthcare institutions.
• Early detection of scoliosis and reduced treatment costs.
• Availability of diagnostic rural areas.

Exploring Retinal Blood Vessel Segmentation with AI: New Insights for Early Disease Detection

S3 Domain: Medicine

Summary: The retina, the eye’s light-sensitive layer, holds valuable clues about various health conditions, including diabetes, hypertension, and cardiovascular diseases. Early and accurate detection of these diseases through retinal imaging can save lives and improve patient outcomes. This paper focuses on using convolutional neural networks (CNNs) for segmenting blood vessels in retinal images, presenting a pixel-wise classification method to aid automated diagnosis.

Challenge: Segmenting retinal blood vessels is challenging due to their complexity and the necessity for high precision in medical diagnosis. Manual segmentation is labor-intensive and prone to errors, making an automated approach highly desirable. However, current models often struggle with accuracy, processing time, and the handling of diverse image qualities.

Proposed Solution: The research introduces a CNN-based approach that processes small image blocks pixels to classify the central pixel, thus effectively segmenting blood vessels. By leveraging deep learning, this method achieves over 95% accuracy, demonstrating a high degree of reliability and robustness across different image qualities. Using high-performance computing (HPC), the solution further optimizes segmentation time through parallel processing.

Business Impact: This AI-based segmentation tool could streamline retinal analysis in clinical settings, reducing the workload on medical professionals and enabling quicker diagnoses. By improving diagnostic efficiency, it supports better resource allocation and enhances the scalability of retinal health monitoring.

Potential Benefits

• Faster and more accurate retinal vessel segmentation for early disease detection
• Reduced manual effort and error rates in clinical analysis
• Increased diagnostic accuracy in diabetic retinopathy, glaucoma, and cardiovascular diseases
• Potential scalability for broader clinical applications through HPC parallel processing
• Improved accessibility of automated diagnostics in underserved regions

Assessing Cardiovascular Risk: A Machine Learning Approach

S3 Domain: Medicine

Summary: Cardiovascular diseases (CVD) remain the leading cause of mortality worldwide, driven by factors like lifestyle choices, genetics, and age. Early risk assessment is crucial, allowing preventive measures to be implemented before critical health issues arise. This study explores using machine learning (ML) models to predict 10-year cardiovascular risk, aiming for a more precise and individualized approach compared to traditional scoring methods.

Challenge: Traditional risk assessment methods, such as the Framingham Risk Score and others, often lack accuracy across diverse populations. These tools sometimes fall short in generalizability, as they were initially calibrated on specific demographic groups. This creates a need for adaptable, data-driven approaches that can be universally applicable while improving predictive accuracy.

Proposed Solution: The study proposes using several machine learning algorithms, including Naive Bayes, Random Forest, K-Nearest Neighbors, Logistic Regression, and Support Vector Machines, to predict cardiovascular risk. The data consists of demographic and health indicators, like age, blood pressure, smoking status, and cholesterol levels, tailored to categorize risk into four levels: low, medium, high, and very high. Through a two-phase experiment, models were trained, validated, and refined using techniques like oversampling to balance the dataset and improve model performance.

Business Impact: Integrating ML-driven risk assessment tools into healthcare could enable more targeted prevention strategies, improving patient outcomes and potentially reducing healthcare costs. Enhanced prediction capabilities allow healthcare providers to tailor intervention plans to each patient’s specific risk profile, supporting better resource allocation and proactive care.

Potential Benefits

• Enhanced Accuracy: ML models can achieve higher accuracy by learning from diverse datasets, reducing misclassification in risk levels.
• Scalability: Once trained, these models can be implemented across various healthcare settings without needing recalibration for each population.
• Preventive Care: Improved risk prediction allows for early lifestyle or medical interventions, potentially preventing 90% of CVD cases.
• Data-Driven Decisions: ML models offer a transparent approach to factor prioritization, enabling physicians to make more informed treatment decisions.

Edge AI in Agriculture: Advancements in Poultry Farming

S3 Domain: Agriculture

Summary: The rapid growth in global demand for animal protein, especially poultry, is pushing the agriculture industry to adopt innovative technologies. One promising approach explored in recent research is the use of Edge AI to enhance poultry farm management. By integrating deep learning models into edge devices, such as IoT-enabled cameras, poultry farmers can monitor flock health and optimize production processes efficiently.

Challenge: Poultry farms face increasing demands for efficient production, along with the challenge of maintaining animal welfare and environmental standards. Essential tasks like monitoring flock health, ensuring proper sanitation, and managing resources require constant oversight, which is time-intensive and complex. Traditional monitoring systems are often unable to keep up with the scale and demands of modern poultry production.

Proposed Solution: The research proposes an Edge AI-based solution using the YOLOv7 Tiny model on a compact NVIDIA Jetson Nano device. This setup can detect and monitor poultry in real-time using camera feeds, enabling automated tracking of bird health and environmental conditions. The use of data augmentation and advanced computer vision techniques allows the model to accurately detect objects (e.g., birds) in resource-constrained edge environments.

Business Impact: Integrating Edge AI in poultry farming can significantly streamline operations, reducing manual monitoring requirements and enhancing response times. This solution also improves decision-making by providing real-time data on animal welfare, helping farms optimize resource usage while maintaining high production standards.

Potential Benefits

• Enhanced Efficiency: Automated monitoring reduces labor costs and optimizes farm operations.
• Improved Animal Welfare: Real-time health tracking allows early detection of issues, improving animal care.
• Environmental Benefits: Efficient resource management (feed, water, energy) reduces environmental impact.
• Scalability: Edge AI systems can be scaled to other areas of agriculture, supporting broader digital transformation.

Leveraging Edge AI for Enhanced COVID-19 Prevention Through Mask Detection

S3 Domain: Health and Tourism

Summary: In response to the COVID-19 pandemic, this research investigates the application of Edge AI for real-time detection of face masks, aiming to enhance health safety in public and healthcare settings. The study utilizes machine learning models deployed on edge devices like the Nvidia Jetson Nano, enabling rapid local processing without reliance on remote servers. The approach not only reduces latency but also lowers bandwidth usage by processing data at the device level.

Challenge: With a global rise in data generated by IoT devices, traditional cloud-based systems struggle with latency and bandwidth constraints, especially for real-time applications such as mask detection. Processing data on distant servers can result in delays that hinder immediate decision-making, critical in health monitoring and infectious disease prevention.

Proposed Solution: The study proposes a system where mask detection is performed directly on an edge device, the Jetson Nano. Using a Faster R-CNN model, trained with a large dataset, the solution can identify people wearing or not wearing masks in real-time. This configuration minimizes response times and enables the system to function independently of internet connectivity, making it suitable for diverse environments, from hospitals to shopping malls.

Business Impact: Implementing Edge AI for mask detection has the potential to streamline operations by automating surveillance tasks, reducing the need for manual monitoring. This can help organizations comply with health regulations while maintaining a safe environment for employees and customers, thus building public trust and reducing pandemic-related disruptions.

Potential Benefits:

• Real-time Monitoring: Instant mask detection enables rapid action in high-traffic areas.
• Reduced Cloud Dependence: Local data processing reduces bandwidth and reliance on internet connectivity.
• Privacy Assurance: On-device processing enhances data security by limiting the need to transfer sensitive data to cloud servers.
• Scalability: The Edge AI model can be adapted to various IoT setups, from airports to educational institutions.

Unlocking Disease Prediction with AI: A Dive into Machine Learning Models

S3 Domain: Health and Tourism

Summary: In today’s fast-paced world, the demand for efficient healthcare solutions is higher than ever. A recent study explored the power of AI in disease diagnosis by comparing the performance of three machine learning models: Support Vector Machine (SVM), Naive Bayes (NB), and Random Forest (RF). Using data on symptoms for a variety of diseases, the study aimed to identify the best-performing model and enhance it through fine-tuning to achieve optimal diagnostic accuracy.

Challenge: With an anticipated shortage of 12.9 million healthcare professionals by 2035, the healthcare industry faces significant challenges. The COVID-19 pandemic underscored the need for rapid and accurate diagnostic tools that can assist overburdened healthcare workers. Predictive models that can diagnose based on symptoms could provide essential support in managing and diagnosing diseases effectively.

Proposed Solution: The study implemented three popular algorithms—SVM, Naive Bayes, and Random Forest—using a dataset of documented symptoms across 42 diseases. The Random Forest model outperformed the others, achieving 87% accuracy. By fine-tuning the Random Forest model with techniques like Grid Search CV, the researchers were able to improve accuracy further, reaching a final performance of 90%.

Business Impact: This AI-driven approach has substantial potential to transform healthcare by automating disease prediction, reducing diagnostic errors, and supporting healthcare professionals in clinical decision-making. Machine learning models can analyze complex data swiftly and accurately, offering a way to bridge gaps in medical expertise and resources.

Potential Benefits

• Increased Diagnostic Accuracy: Machine learning models provide reliable, data-driven diagnoses.
• Resource Efficiency: AI can reduce the burden on healthcare professionals, especially in resource-constrained settings.
• Cost Savings: By identifying diseases early, AI tools can reduce the need for costly advanced treatments.
• Scalability: The solution can be scaled across different healthcare systems, providing widespread benefits.
• Enhanced Patient Outcomes: Early and accurate diagnosis improves the chances of successful treatment and patient recovery.

Breast Cancer Detection: AI and HPC for Faster, More Accurate Diagnostics

S3 Domain: Medicine

Summary: The research explores the application of artificial intelligence (AI), specifically high-performance computing (HPC) combined with convolutional neural networks (CNNs), to detect breast cancer in mammographic images. By implementing models like ResNet152 and DenseNet121, the study aims to enhance the accuracy and efficiency of breast cancer diagnostics, providing early detection and potentially improving treatment outcomes​.

Challenge: Breast cancer detection remains a diagnostic challenge due to the variability in image quality and the subtle nature of malignant changes in mammograms. Traditional diagnostic methods can be time-consuming and sometimes imprecise, which impacts early detection rates critical for successful treatment. Additionally, large datasets needed for training robust AI models are challenging to process without substantial computational resources.

Proposed Solution: The study proposes utilizing advanced CNN architectures, specifically ResNet and DenseNet, in conjunction with HPC resources. This setup allows for the efficient processing of large datasets and the training of deep learning models capable of identifying malignant formations with a high degree of accuracy (AUC score > 0.9). The HPC infrastructure enables simultaneous processing of vast image data, optimizing the speed and accuracy of the model’s performance​.

Business Impact: Implementing AI driven diagnostics in breast cancer detection can improve operational efficiency in medical facilities by reducing diagnostic times and enhancing accuracy. This advancement also has the potential to lower healthcare costs associated with cancer treatment by enabling early-stage detection, thus reducing the need for more invasive procedures. Enhanced diagnostic reliability builds trust in AI-assisted healthcare solutions, paving the way for broader AI adoption in medical diagnostics​.

Potential Benefits

• Increased Diagnostic Accuracy: Achieves high AUC and precision scores, indicating reliable performance in clinical settings.
• Faster Diagnosis: Accelerates the diagnostic process, potentially reducing patient wait times and anxiety.
• Enhanced Early Detection: Supports early intervention, which is crucial for improving breast cancer treatment outcomes.
• Resource Efficiency: HPC reduces the time and computational load required, making it feasible to implement on a larger scale.
• Broader AI Adoption: Successful implementation in breast cancer diagnostics encourages the integration of AI in other areas of healthcare.

Title: Leveraging HPC and AI for Enhanced Diagnosis in Medicine

S3 Domain: Medicine

Summary: The increasing prevalence of thyroid autoimmune diseases has posed challenges in accurate diagnosis due to symptom similarity with other conditions. This research explores how machine learning (ML) algorithms, leveraging high-performance computing (HPC) infrastructure, can aid in the precise classification of thyroid conditions based on patient hormone data, potentially reducing diagnostic ambiguity and improving patient outcomes.

Challenge: Diagnosing thyroid diseases is complex due to overlapping symptoms with other medical conditions, such as depression. Conventional diagnostic methods often fail to differentiate these diseases accurately, which can lead to delayed or incorrect treatments. This study highlights the need for reliable diagnostic support tools that can assist healthcare professionals in identifying patterns within patient data.

Proposed Solution: This study applied various ML models, including Logistic Regression, Random Forest Classifier, Naive Bayes, and Support Vector Machines (SVM), to classify patients as either having a thyroid condition or being healthy. Using a dataset with over 2000 patient records, these algorithms analyzed hormone levels, such as TSH, FT4, TT3, and SHBG, to make classifications. Evaluation metrics—accuracy, precision, recall, and F1 score—indicated that some models were highly effective but also presented overfitting issues due to “perfect” data conditions.

Business Impact: The implementation of HPC-driven AI solutions in healthcare, specifically for thyroid disease diagnosis, could streamline and improve diagnostic processes, especially in high-demand medical facilities. By integrating AI models into healthcare systems, medical professionals can access more reliable diagnostic tools, reducing the time and resources spent on differential diagnosis. This, in turn, enhances patient satisfaction and confidence in healthcare services.

Potential Benefits

• Improved diagnostic accuracy and speed for thyroid conditions
• Reduction in diagnostic ambiguity, leading to more timely treatments
• Enhanced patient outcomes through early and accurate diagnosis
• Potential to apply similar AI models to other complex medical diagnoses
• Opportunity for healthcare facilities to adopt more data-driven approaches

High-Performance Computing and AI in the Energy Sector

S3 Domain: Energy

Summary: This researchpresents a comprehensive review of the application of deep learning models within high-performance computing (HPC) environments to optimize energy sector processes. The research focuses on how advanced models such as Artificial Neural Networks (ANNs), Long Short-Term Memory (LSTM) networks, and hybrid CNN-LSTM models can enhance the accuracy of day-ahead electricity price forecasting.

Challenge: The energy sector faces complex challenges, including the integration of renewable energy sources and the need for accurate demand forecasting. Traditional models struggle with data complexity and fail to achieve the necessary forecasting precision, leading to inefficiencies and increased costs in energy management.

Proposed Solution: This study evaluates the use of HPC resources to implement deep learning models that handle large-scale data, enhancing forecasting accuracy for electricity prices and enabling better decision-making. Using various case studies, the thesis demonstrates the performance of different models in real-world energy markets. This review study underscores the transformative potential of HPC-powered AI in addressing energy sector challenges and improving sustainability. Next step will be experimenting with different datasets openly available to researchers.

Business Impact: Applying deep learning models on HPC infrastructures can lead to more stable energy pricing and improved energy management strategies, crucial for both suppliers and policy makers. Enhanced forecasting accuracy directly impacts operational efficiency, aiding grid operators in optimizing energy distribution and reducing the environmental footprint.

Potential Benefits:

• Improved forecast accuracy for electricity prices, aiding market stability.
• Enhanced integration of renewable energy sources with efficient demand management.
• Reduced operational costs and improved resource allocation in the energy sector.
• Proactive grid management, reducing outages and enhancing resilience.
• Better-informed policy decisions through reliable predictive insights.

Generative AI in Poultry Farming: Dataset and Model Training for Chicken Detection

S3 Domain: Agriculture

Summary: To meet the ever-increasing demands for efficient poultry farming, new technologies are revolutionizing how we manage farm operations. Recent research highlights the use of Generative AI to create synthetic datasets, which are instrumental for training advanced machine learning models for chicken detection. By harnessing synthetic data, poultry farmers and agritech solutions can now overcome data scarcity, facilitating the development of accurate and reliable models that support efficient and automated farm management.

Challenge: One of the primary challenges in poultry farming is obtaining large, high-quality datasets to train machine learning models. Capturing comprehensive datasets with variations in lighting, environment, and flock behavior is often resource-intensive, time-consuming, and costly. These limitations hinder the development of robust computer vision models needed for effective monitoring and management of poultry, impacting efficiency and scalability.

Proposed Solution: The research proposes utilizing Generative AI to create artificial datasets for training chicken detection models. By generating realistic synthetic images of chickens in various environments and lighting conditions, this approach enables a richer and more diverse dataset than would be feasible through traditional data collection. These synthetic datasets are used to train machine learning models, which are then evaluated on real-life images to test their accuracy and robustness.

Business Impact: Using Generative AI to create datasets for chicken detection significantly reduces data collection costs and accelerates the model training process. This advancement allows poultry farms to deploy effective, data-driven solutions without the heavy reliance on extensive real-world data. As a result, poultry operations can improve monitoring accuracy, optimize resource use, and streamline overall farm productivity.

Potential Benefits

• Cost Efficiency: Reduces data collection expenses by using synthetic images to supplement real-world data.
• Scalability: Enables faster training of detection models, making it easier to implement solutions across large-scale operations.
• Enhanced Accuracy: Creates diverse datasets that improve model robustness across varied real-life conditions.
• Accelerated Innovation: Frees up resources for research and development in other critical areas of farm management.