Dataingenjör på Utexpo

AI-driven classification of prostate cancer – using 3D-MRI scans

• Participants: Jasmine Alani and Ismail Demir.
  

Prostate cancer is one of the most common types of cancer among men and requires early and reliable diagnosis to improve patient outcomes. Currently, diagnosis largely relies on the radiologist’s interpretation of multiparametric magnetic resonance imaging (mpMRI). However, studies show that radiologists agree in less than 50 % of cases, which means that patients are at risk of receiving incorrect or delayed diagnoses. This thesis addresses the issue by developing a custom CNN model while comparing it with pre-trained AI-based systems that can automatically identify mpMRI images with high precision and sensitivity. The target audience is healthcare professionals, particularly radiologists, who can benefit from reliable AI support tools to reduce subjective uncertainty in diagnostics, improve decisio-making, and ultimately enhance the quality of patient care.

Comparative Analysis of Unsupervised and Supervised Topic Models in IT Ticket Classification

• Participant: Ibrahim Halil Demir.
  
• Collaborator: Västerviks kommuns IT-enhet.

Public sector IT departments handle a high volume of support tickets every day, each of which must be categorized to reach the correct team. Today, this is often done manually, which is time-consuming and inconsistent. This project explores how machine learning can automate the classification process, even when no labeled training data is available.

Three models were compared: a supervised BERT-based classifier, and two unsupervised models: BERTopic and LDA. The models were trained and evaluated using a synthetic dataset designed to mimic real IT support cases from Västervik Municipality, based on ten predefined categories.

The results show that the BERT model achieved the highest macro F1-score (83 %). However, a key insight is that BERTopic, despite not using any labels, reached 74 % – just 9 % lower. This suggests that BERTopic is a strong alternative in environments where labeled data is unavailable or costly. The project demonstrates that machine learning can improve efficiency in IT support while enabling scalable and GDPR-compliant automation for the public sector.

Evaluating compression algorithms for Linux hibernate/resume performance

• Participants: Richard Lindwall and Albert Ågren.
• Collaborator: Volvo Cars Corporation.

This thesis evaluates how various compression algorithms affect Linux suspend-to-disk (hibernate) and resume performance, aiming to improve boot times for vehicle ECUs. The project evaluates the algorithms LZ4, LZO, and ZSTD, as well as disabling hibernation image compression, by using benchmark tests and actual implementations, such as QEMU virtualization and a stationary computer setup. Each algorithm was assessed for compression and decompression speed, compression ratio, and overall hibernate/resume performance. A custom Linux kernel was modified to support ZSTD, as it is not available by default. Results indicate that while ZSTD offers the best compression ratio by about 25 %, it performs worse than LZ4 and LZO during both hibernation and resume when it comes to speed – hibernating half as fast and resuming 25 % slower than LZ4. Although LZ4 was the best among compression methods, omitting compression entirely yielded the fastest suspend/resume cycle. Additional tests evaluated the impact of page utilization, for example how much each chunk of allocated memory is used, on image size and hibernate/resume performance, where a higher percentage of utilization led to a larger compressed image size and slightly lower compression performance.

Optimizing chemical risk documentation with LLMs

• Deltagare: Alaa Mouselli och Faraj Hamedy.
• Samarbetspartner: cDOC kemikaliedokumentation.

Denna studie presenterar en innovativ lösning baserad på artificiell intellgens, som i allt större utsträckning används för att verifiera data i databaser. Systemet syftar till att upptäcka fel, inklusive dubbletter, tomma fält eller felaktiga beskrivningar av kemiska produkter som lagras i databasen. Projektet är i första hand utformat för att stödja säkerhetsbedömningar inom industrin genom att automatisera dataverifiering via jämförelser med externa källor med hjälp av språkmodellen GPT-4 Turbo. Simulerade professionella tester visade mätbara förbättringar i feldetektion och riskprioritering, där inlärningskurvorna steg från 30 % till 60 % noggrannhet efter flera justeringar av promptarna. Projektet tar upp utmaningar som LLM-hallucinationer, tillförlitlig återkoppling och krav vid säkerhetskritiska tillämpningar, och erbjuder en skalbar grund för semiautomatiserad validering av kemisk dokumentation i industriella miljöer.
Sammanfattningsvis visar detta att artificiell intelligens kan integreras för att förbättra datakvaliteten och därmed minska den mänskliga arbetsbördan i miljöer som kräver både noggrannhet och snabbhet.
Nyckelord: Feldetektion, automation, databas, prompt, återkopplingsslinga, kontaktsystem, Large Language Model (LLM).

Remote Operation of Electrically Power-Assisted Cycles (EPAC)

• Deltagare: Mehari Goitom Tewelde och Najib Abshir Dirir.
  

Detta projekt handlar om att utveckla ett innovativt system för fjärrövervakning och kommunikation för elassisterade cyklar (EPACs) med hjälp av V2X-teknik (Vehicle-to-Everything). Syftet är att skapa en lösning som gör det möjligt att i realtid samla in och skicka data mellan en EPAC och en extern mottagare, exempelvis en server, dator eller mobilapplikation. Genom att använda en kombination av CAN-buss, Wi-Fi och mikrokontrollersystem, som ESP32 och Raspberry Pi Pico, överförs information såsom hastighet, batterinivå, motorstatus och positionsdata trådlöst.

Systemet bygger på ett säkert och effektivt kommunikationsprotokoll och kan anpassas för olika användningsområden, som till exempel delningscyklar, logistik, forskning eller förebyggande underhåll. Projektet innefattar hårdvaruintegration, programmering i C och Python, implementation av realtidsövervakning och dataloggning. Det färdiga systemet kan också utvidgas till att stödja fjärrstyrning i kontrollerade miljöer.

Genom att koppla ihop fordon och infrastruktur på ett smart sätt bidrar projektet till framtidens hållbara transportsystem och öppnar upp för nya lösningar inom mobilitet, trafiksäkerhet och smarta städer.

Usefulness of Synthetic Data in Ocular Recognition

• Participants: Adrian Sterner and Filip Thunberg.
  

Ocular recognition plays a crucial role in scenarios where masks or partial faces limit the use of face recognition. However, privacy regulations and restrictions have hindered access to large-scale public databases necessary for training and evaluating recognition models. This thesis investigates whether synthetic face data can substitute real images while retaining the detail required for ocular recognition.

Visualization of patient progress

• Participants: Aleksander Jankovic and Ludvig Sanell.
  
• Collaborator: Ensolution.

In this thesis, we have created an interactive follow-up module for the digital healthcare system, Kuben, in collaboration with the company Ensolution. The goal of the module is to help healthcare staff track how patients are doing over time. It uses visual tools, such as graphs and charts to showcase key numbers.

The system is built using ReactJS, with Recoil for managing the state and Recharts for the graphs. It then connects to an existing C# .NET backend utilizing a RESTful api approach. The project has worked in an iterative way, receiving feedback from healthcare staff, managers, and developers during the process. This has led to the final result being a responsive interface that lets users compare patient data in multi-period views and filter it by region, among other things.

During user testing, the module proved easy to use, clear, and efficient. Overall, the project shows how visual tools can support clinical decisions and reduce admin work, helping to move toward more data-based and sustainable healthcare.