Mathematik  |  Informatik

 

Mikkel Lüscher, 2006 | Erlinsbach, AG

 

Brain tumor segmentation is a manual, time-consuming process subject to inter-observer variability. This project develops NOSA, an AI-driven segmentation tool built on a systematically improved U-Net architecture, achieving a mean Whole Tumor Dice of 88.29% and IoU of 80.41% on the BraTS 2023 GLI dataset, within two percent of the challenge-winning model. The research includes an interview with a neuroradiologist on AI in radiology.

Introduction

Can an AI-driven diagnostic tool be developed and improved toward clinical relevance through controlled architectural changes?

Methods

NOSAv2 was a proof-of-concept 2D U-Net with 4 pooling layers, trained on the Jun Cheng dataset of 3064 contrast-enhanced MRI images from 233 patients, using combined Dice and binary cross-entropy loss with an Adam optimizer. A statistical heatmap filtered false positives in the skull and eye regions. NOSAv2 included a GUI built with PyQt5.
NOSAv3 was trained on the BraTS 2023 GLI dataset, a clinical-grade benchmark of 1251 multimodal MRI cases, using a computer built through crowdfunding featuring an NVIDIA RTX 5090. All versions were optimized for Whole Tumor segmentation. Earlier versions were trained on the first fold for direct comparison; NOSAv3.3 was trained on all five folds.
NOSAv3 established the 3D baseline. NOSAv3.1 introduced the core architecture: an asymmetric 3D U-Net with anisotropic kernels, InstanceNorm, LeakyReLU, Kaiming initialization, and tumor-biased patch sampling. NOSAv3.2 introduced residual blocks. NOSAv3.2.1 added deep supervision and a warmup-to-cosine schedule. NOSAv3.3 added multi-scale attention at the input, encoder, bottleneck, and decoder levels.
The GUI for NOSAv3.3 displays inference speed, IoU, and Hausdorff distance per prediction, and includes a rule-based tumor detection module. The research was concluded with an interview with Dr. Diepers, Deputy Head of Neuroradiology at the Kantonsspital Aarau.

Results

NOSAv2 achieved an IoU of 32.1%, significantly below state-of-the-art models. The NOSAv3 series showed systematic improvement: peak validation IoU during training reached 83.12% for v3.1, 83.96% for v3.2, 84.05% for v3.2.1, and 85.02% for v3.3. The final model, NOSAv3.3, was evaluated using 5-fold cross-validation, yielding a mean Dice score of 88.29%, a mean IoU of 80.41%, a mean recall of 89.75%, a mean HD95 of 5.68 mm, and a mean inference time of 0.87 seconds per case. In practical terms, this means the model identifies approximately 88 out of every 100 tumor voxels correctly, and its boundary deviates by less than 6 mm from the expert annotation. Dice and IoU remained within 1.0% and 1.2% respectively across all five folds, with recall showing slightly more variation of 4.3%, indicating strong generalizability. NOSAv3.3 achieves a Whole Tumor Dice only 1.76% below the winning solution of the BraTS 2023 Adult Glioma challenge, and a notably better HD95 of 5.68 mm compared to 14.94 mm. This comparison should be interpreted with care, as the winning model was evaluated on a hidden test set, while NOSAv3.3 results come from 5-fold cross-validation on the training data.
The interview confirmed AI tools are already used clinically for hemorrhage detection and perfusion mapping, but always under radiologist supervision.

Discussion

The results confirm the research goal has been reached. Each change from v3.1 to v3.3 produced a measurable IoU improvement, and NOSAv3.3 reaches 80.41% IoU, approaching the BraTS 2023 challenge winner. The transition to 3D processing was the single largest driver of improvement. The attention modules in v3.3 yielded only a modest gain, likely due to absent hyperparameter tuning, which was impractical given the 18-hour training time per run. A self-configuring framework such as nnU-Net may have achieved stronger results with less manual effort. Focusing on Whole Tumor segmentation rather than all subregions was the right methodological choice, simplifying the task while remaining sufficient for a detection tool. A key self-critical point is that earlier versions were evaluated on only a single fold, which limits the reliability of inter-version comparisons; full cross-validation for all versions would have been more rigorous but was not feasible given available compute.

Conclusions

NOSAv2 laid the groundwork for the v3 series. NOSAv3.3 achieves 80.41% IoU and demonstrates that a systematically improved U-Net can approach state-of-the-art performance. Next steps include an AI-driven diagnostic module for tumor characterization and generative AI for training data augmentation. The long-term goal is to submit NOSAv3.3 to BraTS 2026.

 

 

Würdigung durch den Experten

Mathieu Guillame-Bert

This research project implements a deep learning model for automated brain tumor segmentation on MRI data. The candidate showed technical proficiency by implementing a U-Net inspired model and testing enhancements like multi-scale attention and data-aware batch generation. The methodology was rigorous, evaluating models on a competition dataset and achieving results within 2% of the winning solution. Beyond ML work, a GUI was developed for clinical evaluation with a neuroradiologist. The project shows a capacity for research, problem solving, and translating theoretical concepts into products.

Prädikat:

Gold

Sonderpreis «European Union Contest for Young Scientists (EUCYS)» gestiftet von der Stiftung Aldo e Cele Daccò

 

 

 

Alte Kantonsschule Aarau
Lehrerin: Martina Vàzquez