We are pleased to announce that Lesław Pawlaczyk (CEO) has written a paper that will be presented at the prestigious international conference ICSENG 2025 (https://icseng.eu//). The paper, titled "Multi-Scale U-Net Segmentation Optimized by CNN-Based Quality Scoring," discusses the modification of vessel segmentation in fundus images using a voting system and random window position generation. This technology is used in our products. The abstract is below:
This paper presents Multi-Scale U-Net Segmentation Optimized by CNN-Based Quality Scoring MUSOCS, a novel framework designed to address key limitations of the traditional U-Net in medical image segmentation. Standard U-Net models struggle with resolution variability and often require patch-based processing, introducing artifacts such as checkerboard effects. MUSOCS mitigates these challenges by combining adaptive multi-scale segmentation with a CNN-based scoring mechanism that automatically selects the optimal resolution for each image. The framework further enhances accuracy by integrating a hybrid strategy of full-image mesh coverage and randomly selected windows ranked by quality scores. We evaluate MUSOCS using models trained on the FIVES dataset and tested on a combined validation set derived from multiple sources, demonstrating improved robustness and scalability. Results show an Average Jaccard Index gain of up to 0.04 compared to classical Sliding Window Approaches, validating the effectiveness of adaptive scale selection. The complete implementation is open source and available on GitHub.
This aligns with our broader strategy of publishing research articles in the fields of ophthalmology and artificial intelligence. We will be announcing additional articles in the coming months.
