The Austrian Society for Artificial Intelligence (ASAI) honors exceptional young researchers for the best diploma or master’s thesis in Artificial Intelligence (AI) with the ASAI Prize. This year, the first prize was awarded to two winners: Bettina Röthlin and Bernhard Steindl from TU Wien Informatics. The ASAI Prize is endowed with €2000.

Bettina Röthlin’s master’s thesis, “Deep Learning-based Identification of Suspicious Areas in Breast MRIs”, explores how deep learning can support breast cancer risk assessment in high-risk women by analyzing breast DCE-MRI scans. Traditional screening methods, even when done annually, can miss early-stage lesions. To address this, Bettina developed a segmentation-based pipeline that identifies areas in MRI scans that are more likely to develop cancer within 6 to 24 months. She compared several deep learning models for their ability to highlight these high-risk regions, with DeepLabv3+ performing best, especially when enhanced with data augmentation and transfer learning. Features extracted from the segmentation probability maps were then used to train a Random Forest model to classify MRI slices as at risk or not. The results show that while traditional segmentation performance was limited, the probability maps were highly informative for risk prediction.

Bernhard Steindl’s master’s thesis, “Evaluating the Fairness of News Recommender Algorithms Within Detected User Communities”, investigates fairness in recommender systems by studying how different user groups receive recommendations. Recommender systems are commonly used to suggest content based on a user’s past behavior. However, these systems can treat certain groups unfairly, especially when the data is imbalanced. Rather than grouping users by demographics, Bernhard identified large user communities based on actual interaction data from the Austrian news platform “DER STANDARD”. User networks were built from actions like article clicks, comments, votes, and followers. Bernhard then applied community detection algorithms to find natural groupings within the network. The study compares how various collaborative filtering algorithms perform across these communities and finds that recommendation quality varies significantly between them. It also reveals that community detection itself can be unstable, with results depending on how the network is built. Overall, the thesis highlights that both the choice of algorithm and how users are grouped can impact fairness in recommendations, pointing to the need for more robust, equitable approaches in recommender system design.