CS MSc Thesis Presentations 10 June 2026

Wednesday, 10 June there will be two master thesis presentations in Computer Science at Lund University, Faculty of Engineering.

The presentation will take place in E:1426 and E:4130 (Lucas).

Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname [dot] lastname [at] cs [dot] lth [dot] se). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions for opponents are found here on the LTH thesis project page.

10:00-11:00 in E:4130 (Lucas)

• Presenters: Adrian Steene, Lucas Wittich
• Title: Breaking Language Barriers in Real Time: On-Device AI-Driven Speech Translation
• Examiner: Xuan-Son Vu
• Supervisors: Pierre Nugues(LTH), Zebastian Karra-Krüger(Axis), Dan Lundgren(Axis)

This thesis investigates the feasibility of running multilingual speech translation locally under real-time and resource-constrained conditions. We implement a cascaded pipeline in which automatic speech recognition first transcribes source-language speech and machine translation then generates target-language text.

The system is evaluated across six European languages through component-level and end-to-end experiments. The ASR study compares multilingual recognition models and optimizes the strongest candidate using ONNX Runtime and dynamic INT8 quantization. The MT study compares multilingual translation models and fine-tunes the selected model to improve weak language directions. The optimized components are then integrated and evaluated as a complete speech translation pipeline.

The results show that accurate real-time ASR is feasible under an 8 GiB memory budget after deployment-oriented optimization. However, end-to-end performance is primarily limited by MT latency, and translation quality degrades when MT receives ASR-generated transcripts. Local multilingual speech translation is therefore technically plausible, but only with careful system configuration.

Here is a link to the popular science summary

11:15-12:00 in E:4130 (Lucas)

• Presenters: Victor Pekkari, Harry Wik
• Title: Cost-Efficient Graph Database-backed Graph Neural Network Pipelines
• Examiner: Jacek Malec
• Supervisors: Xuan-Son Vu (LTH), Brian Shi (Neo4j), Alfred Clemedtson (Neo4j) 

Graph Databases (GDBs) are essential for managing large-scale interconnected data. While GDB-backed Graph Neural Network (GNN) pipelines are feasible, their performance often lag behind in-memory implementations due to I/O bottlenecks and significant data transfer. However, in comparison to in-memory stores, GDB solutions can drastically cut costs due to the decreased amount of RAM needed, making training possible on resource-constrained machines. This thesis implements a GDB-backed GNN training pipeline that is shown to have functional equivalence to the corresponding standard in-memory implementation. By pushing computation directly into the database engine, we minimize data movement during training, leading to increased throughput. Furthermore, we introduce a novel inference approach, executing the full forward pass in the GDB, eliminating data transfer and making small-batch inference faster.

Here is a link to the popular science summary

13:15-14:00 in E:1426 N.B. Change of room

• Presenters: Andreas Ruggieri, Hugo Nilsson
• Title: Rendering Time Analysis of In-Flight Entertainment Systems
• Examiner: Christoph Reichenbach
• Supervisor: Jonas Skeppstedt (LTH)

This thesis investigates rendering performance in Android-based in-flight entertainment (IFE) systems, where efficient rendering is particularly relevant due to the long hardware lifecycles common in the aviation industry. A prototype Android application was developed and benchmarked on real IFE hardware to compare two UI frameworks, the traditional XML-based Android View system and Jetpack Compose, together with the image-loading libraries Glide and Coil. Rendering performance was evaluated using Android Macrobenchmark by measuring Time to Initial Display (TTID) and Time to Full Display (TTFD).

The results showed that the image-loading library had the greatest impact on rendering performance. Configurations using Glide consistently outperformed those using Coil, particularly during repeated rendering of image-heavy views where caching had a greater effect on performance. However, additional experiments indicated that Coil’s performance could be improved through code optimizations. In comparison, the performance differences between XML-based views and Jetpack Compose were smaller overall, although XML-based views achieved slightly lower rendering times in most tests.

Here is a link to the popular science summary

15:15-16:00 in E:4130 (Lucas)

• Presenters: Elias Hirschfeld, Gloria Liu
• Title: Classifying and Explaining Cervical Cytology with Foundation Models
• Examiner: Jacek Malec
• Supervisors: Pierre Nugues (LTH), Fredrik Nilsson (CellaVision)

This thesis investigates the use of vision-language models and vision transformers for cervical cytology image analysis, with a focus on classification and clinically meaningful reasoning. The work evaluates models including MedGemma, Qwen3-VL, EfficientNetV2-S, UNI2-h, SigLIP-2, and MedSigLIP on a four-class Bethesda-inspired severity scale: Negative, Low grade, High grade, and Malignant. Public and private cervical cytology datasets were harmonized, filtered, augmented, and balanced for training and evaluation. Results show that vision transformers and CNN-based models outperform vision-language models for classification, with fine-tuned ViTs achieving the strongest macro-F1 scores. While larger multimodal models showed some ability to describe relevant cytological features, their reasoning was often inconsistent or hallucinated. The findings suggest that current VLMs are not yet reliable for cytology reasoning without stronger domain-specific visual grounding, but medically adapted vision encoders show promise for future diagnostic support systems.

Here is a link to the popular science summary