Cologne AI and Machine Learning Meetup

CAIML #41 is going to happen on March 17, 2026, at ATVANTAGE.

We will have two talks with additional time for networking.

Talk 1: Moritz Wegener (Data Scientist at ATVANTAGE): Nodes, Networks, and Capital Myths: How Graphs Uncover Dubious Payment Networks

This talk explains how heterogeneous entities can be consolidated into meaningful graph-based units—such as individuals, companies, and corporate groups—and how these structures enable the detection of suspicious payment networks. By modeling ownership, control, and transactional relationships as graphs, we demonstrate how network patterns and structural anomalies help identify candidates for fraudulent activity that remain hidden in traditional, flat data representations.

Talk 2: Gergő Szita (Data Scientist): ICG-enhanced FOI of Rheumatoid Arthritis Classification

Rheumatoid arthritis (RA) is a chronic inflammatory disease affecting around 24.5 million people worldwide. Early detection is critical, yet existing diagnostic methods are either invasive, time-consuming, or heavily dependent on expert interpretation. Fluorescence Optical Imaging (FOI) with indocyanine green (ICG), as used in the Xiralite system, offers a fast, safe, and non-invasive way to visualize inflammatory activity in the hands—but its analysis is still largely manual. In this talk, I present a deep learning–based approach to automate RA classification from ICG-enhanced FOI image sequences. The dataset consists of 10,080 images from 28 patients. To standardize and denoise the data, we applied histogram equalization, threshold-based segmentation, and convex hull extraction. To capture both spatial inflammation patterns and their temporal dynamics, we trained a hybrid model combining a 3D convolutional neural network (CNN) with a long short-term memory (LSTM) network. Class imbalance was addressed using data augmentation.
The model achieved an accuracy of 81.8%, with a precision of 85.7%, recall of 75.0%, an F1-score of 80.0%, and a specificity of 88.0%. Results show that temporal information in FOI sequences is a key factor for reliable classification. The talk discusses model design choices, practical challenges with small medical datasets, and how such systems could complement clinical workflows. Future directions include multi-class severity scoring, alternative architectures, and improved tooling for clinical adoption.

We will share more details on the talks and an agenda soon.

See you in March,
Aaqib, Marc & Fabian