Every year, MDSI honors three outstanding papers published in the previous year by groups of MDSI Core Members. For the first time, this year the prizes were awarded in three categories: Societal Impact, Foundational Impact, and Applications.
Each awardee won €500, and all three winners received a certificate from MDSI.
Category A: Societal Impact
Friederike Jungmann, Paul Hager and their co-authors critically evaluate the performance of large language models (LLMs) like ChatGPT in clinical decision-making in their Nature Medicine publication "Evaluation and mitigation of the limitations of large language models in clinical decision-making".
Recognizing the limitations of traditional testing methods, the team around the MDSI Core Member Prof. Marcus Makowski and MDSI Director Prof. Daniel Rückert at TUM Klinikum rechts der Isar developed a realistic testing framework and dataset designed to evaluate the diagnostic capabilities of LLMs in dynamic, information-sparse environments using only real-world patient cases of four abdominal diseases. LLMs were tasked to actively request diagnostic information and to propose diagnosis and treatment plans, mimicking real clinical workflows. The models were evaluated against junior and senior clinicians across hospitals in Germany and the US. Results showed that LLMs consistently underperformed human counterparts, often deviating from evidence-based treatment guidelines. The study also revealed that LLM performance was highly sensitive to both the order and phrasing of inputs, raising concerns about consistency and reliability.
To promote transparency and progress, the authors made their dataset and framework freely available. Their work underscores the importance of rigorous, clinically grounded evaluation before integrating AI into healthcare settings.
Category B: Foundational Impact
In their publication "MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers", Yawar Siddiqui and a team around the MDSI Core Members Prof. Angela Dai and Prof. Matthias Niessner introduced MeshGPT, a new approach for generating triangle meshes. It reflects the compactness typical of artist-created meshes, in contrast to dense triangle meshes extracted by iso-surfacing methods from neural fields. Inspired by recent advances in powerful LLMs, they adopted a sequence-based approach to autoregressively generate triangle meshes as sequences of triangles. First, the researchers had to learn a vocabulary of latent quantized embeddings, using graph convolutions, which inform these embeddings of the local mesh geometry and topology. These embeddings are sequenced and decoded into triangles by a decoder, ensuring that they can effectively reconstruct the mesh. A transformer is then trained on this learned vocabulary to predict the index of the next embedding given previous embeddings. Once trained, the model can be autoregressively sampled to generate new triangle meshes, directly generating compact meshes with sharp edges, more closely imitating the efficient triangulation patterns of human-crafted meshes.
MeshGPT demonstrated a notable improvement over state-of-the-art mesh generation methods, with a 9% increase in shape coverage and a 30-point enhancement in FID scores across various categories.
The publication was recognized as a "Highlight" by the Conference on Computer Vision and Pattern Recognition (CVPR).
Category C: Applications
The publication of Fabian Pfitzner, MDSI Advisory Board Member Dr.-Ing. Alexander Braun, and MDSI Core Member Prof. André Borrmann, "From data to knowledge: Construction process analysis through continuous image capturing, object detection, and knowledge graph creation", introduced a data-driven pipeline that transforms raw construction site data into structured, actionable knowledge. It combines computer vision, geometric projection, and knowledge graphs to create a time- and location-aware digital twin of construction progress. High-resolution images from crane-mounted cameras are analyzed using deep learning to detect entities like workers, vehicles, and building elements. These detections are projected onto floor plans and linked in a labeled property graph, enabling real-time monitoring of site activity. The system was validated across four construction sites over 1.5 years, generating millions of images and graph relationships.
Their approach allowed precise tracking of productivity, identification of bottlenecks, and querying of deviations from planned schedules. The integrated pipeline bridges unstructured visual data with structured process models, offering scalable solutions for construction and other physical domains. Overall, the work demonstrated how AI can revolutionize monitoring in traditionally low-tech industries by enabling transparent, performance-driven decision-making.
Links:
Category A: Paul Hager, Friederike Jungmann et al., "Evaluation and mitigation of the limitations of large language models in clinical decision-making", Nature Medicine 30, 2613–2622 (2024).
Category B: Yawar Siddiqui et al. "MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers", CVPR 2024.
Category C: Fabian Pfitzner et al. "From data to knowledge: Construction process analysis through continuous image capturing, object detection, and knowledge graph creation", Automation in Construction 164(9), 105451, 2024.