Index

Radiology Report Classification

Evaluate few-shot detection on VinDR-CXR

Accurate localization of thoracic abnormalities in chest X-ray images remains a major challenge due to the limited
availability of large-scale, finely annotated datasets. Few-shot learning has recently emerged as a promising strategy to
address this problem by enabling models to generalize to unseen categories with only a small number of labeled
examples. In this work, we propose an improved few-shot localization approach for VinDr-CXR images by leveraging
the DINO-DETR model, a transformer-based detection framework with self-supervised pretraining. Our method
adapts DINO-DETR to the few-shot setting through task specific fine-tuning and optimization strategies designed to
improve feature alignment between support and query samples. Experimental results demonstrate
that the proposed method achieves competitive localization accuracy compared to baseline approaches, while reducing
the reliance on large annotated datasets. Although certain predictions remain imperfect, particularly in cases with subtle
or overlapping pathologies, the approach shows clear potential for scaling to broader medical imaging applications.
This study highlights both the opportunities and limitations of applying state-of-the-art transformer-based detection
architectures to few-shot medical image localization and suggests directions for future improvements, such as data
augmentation and cross-domain pretraining.

Deep Learning-Based Classification and Explainability of Cytomegalovirus Encephalitis in Longitudinal MRI Data

This Master Project focuses on developing and evaluating advanced deep learning methodologies for the automated detection and classification of Cytomegalovirus (CMV)-induced encephalitis using clinical Magnetic Resonance Imaging (MRI) data.

Motivation and Goal

CMV encephalitis is a challenging condition to diagnose, and advanced, non-invasive computational methods are required to assist clinical decision-making. The primary goal is to leverage the temporal and multi-modal information within longitudinal MR scans to classify the presence or stage of inflammation (encephalitis).

Data and Scope

The project utilizes a unique, high-quality, pre-selected longitudinal dataset comprising MRI scans from approximately 300 patients, with an average of six scan time points and multiple MR sequences available per visit.

Key Tasks and Research Questions

  1. Literature Review: Conduct a targeted literature review of similar projects focused on MR classification, specifically those dealing with longitudinal data (e.g., the BraTS Challenge: Predicting the Tumor Response During Therapy) and methods for medical image classification and prediction.
  2. Model Development: Adapt and implement state-of-the-art deep learning architectures (e.g., 3D Convolutional Neural Networks, Recurrent Neural Networks, or hybrid models) suitable for processing longitudinal and multi-sequence volumetric data.
  3. Explainability (XAI): A critical component of the project is the integration of Explainable AI techniques (e.g., Grad-CAM, Saliency Mapping, or LRP). The student will implement and evaluate these methods to highlight which anatomical regions or temporal patterns contribute most significantly to the model’s classification decision, thereby increasing clinical trust and interpretability.
  4. Evaluation: The core research question is whether robust classification performance can be achieved on the available data, considering potential constraints such as fewer advanced sequences or less acute disease stages compared to published reference literature. Performance will be measured using metrics like AUC, sensitivity, and specificity.

Generalization of Self-Supervised Vision Models in Image Retrieval

This project investigates the generalization capabilities of prominent self-supervised vision models, such as DINOv2, CLIP, and MoCo, when applied to image retrieval tasks across diverse visual domains.

Overview

Self-supervised learning (SSL) models are increasingly important for creating robust visual features. However, their performance often degrades when transferring from standard natural image datasets (like ImageNet) to more specialized domains.

Our study systematically benchmarks these models on three distinct domains:

1. Natural Images (Baseline)

2. Scene-Centric Images (e.g., Places365)

3. Artistic Images (e.g., ArtPlaces)

We compare the models’ out-of-the-box generalization and analyze the impact of finetuning on their domain adaptation, providing crucial insights into the stability and robustness of the learned representations for practical retrieval applications.

AI for Inflammatory Skin Pathology: Psoriasis and Eczema Classification in Whole-Slide Images

Exploring foundation models for high-resolution whole-slide image classification, with a focus on interpretable predictions through attention maps that align with medically relevant regions.

Evaluating Time-Frequency Representations for Intelligent Fault Analysis in Power System Protection

This project investigates how different time-frequency representations, including spectrograms, wavelets, scalograms, and Gramian Angular Fields, affect deep learning performance in fault analysis tasks. Using high-resolution current and voltage signals, it benchmarks various representation strategies across multiple neural network architectures. The goal is to determine which transformations capture fault dynamics most effectively and enable robust model generalization. The findings will guide the design of future machine learning based protection systems for resilient and data-driven power grids.

Multi-Task Learning for Integrated Fault Analysis in Power System Protection

Modern electric power systems require rapid and reliable fault analysis to ensure grid stability amid increasing renewable integration. This project explores multi-task learning as a unified framework for simultaneously detecting, classifying, and localizing faults in transmission networks. By sharing representations across tasks, the model aims to reduce redundancy and enhance generalization compared to traditional single-task approaches. The results will contribute to the development of scalable, data-driven protection schemes for future intelligent power grids.

Reproducible Reinforcement Learning on a Real-World Power Grid Control Problem

This is a project only (10 ECTS) focused on reproducible reinforcement learning and paper-driven implementation.
You will re-implement an IEEE-published RL method [1] and evaluate it on a realistic, safety-critical control problem.

The application domain is power grids, used purely as a real-world benchmark for reinforcement learning.
No prior power-systems background is required.

  • Implement a Q-learning-based control method from a research paper (state/action design, reward shaping, constraints).
  • Validate the implementation on a benchmark setup (reproducibility, metrics, sanity checks).
  • Apply the method to real data from a 20 kV distribution grid.
  • Optional: extend the same RL framework towards distance protection (concept + first prototype, time permitting).

Who should apply

  • Computer science or related background.
  • Good Python skills (NumPy/Pandas, Git).
  • Basic knowledge of machine learning or reinforcement learning.
  • Interest in implementing and evaluating methods from scientific papers.
  • Able to attend the weekly in-person meeting in Erlangen (Mondays, 14:00).

Apply

Send one PDF to julian.oelhaf@fau.de with the subject:

"Application | Project (10 ECTS) | Reproducible RL on Power Grids | <Your Full Name>"

Email body (max. 200 words): Short motivation and your earliest state date.

Attach as one PDF: CV, transcript (dated), optional code links.

📌 Incomplete applications will not be considered.

References

[1] H. C. Kılıçkıran, B. Kekezoglu, and N. G. Paterakis, Reinforcement Learning for Optimal Protection Coordination, IEEE SEST, 2018. DOI
[2] D. Wu, X. Zheng, D. Kalathil, and L. Xie, Nested Reinforcement Learning-Based Control for Protective Relays in Power Distribution Systems, IEEE CDC, 2019. DOI
[3] D. Wu, D. Kalathil, M. M. Begovic, K. Q. Ding, and L. Xie, Deep Reinforcement Learning-Based Robust Protection in DER-Rich Distribution Grids, IEEE Open Access Journal of Power and Energy, 2022. DOI

A Resource-Efficient AC Power Flow Prediction Framework using Physics-Informed GNNs and RL-Based Model Compression

1. Motivation

Modern power grids require accurate, real-time AC power flow prediction to ensure secure and efficient operation. Graph Neural Networks (GNNs) are promising due to their ability to model the grid’s topological and nonlinear properties. However, standard GNNs are often too large for edge deployment, and naïve compression can lead to physically infeasible predictions. There is a pressing need for compression techniques that preserve physical accuracy.

2. Objective

This project aims to develop a two-phase framework:
1. Physics-Informed GNN: Predict voltage magnitudes and phase angles from power grid snapshots using AC power flow laws.
2. RL-Guided Compression: Learn to prune and quantize the model efficiently while preserving physical feasibility.

3D detection of Abdominal Trauma in CT images using cross-attention