Основные понятия
The author proposes an architecture that leverages high-resolution ground truths to improve prediction quality while using low-resolution inputs efficiently.
Аннотация
The content discusses the challenges of deploying neural networks in medical devices due to hardware limitations. It introduces a novel architecture that enhances prediction quality by utilizing high-resolution ground truths with low-resolution inputs. The proposed model shows significant improvements in semantic segmentation for cancer detection in MRI images, outperforming existing state-of-the-art frameworks. Extensive experiments on different datasets validate the effectiveness of the approach, showcasing better prediction quality and reduced computational complexity.
Статистика
Using lower-resolution input leads to a significant reduction in computing and memory requirements.
The proposed model improves prediction quality by 5.5% with less than 200 additional parameters.
Our architecture reaches higher prediction scores with less model complexity compared to standard U-Net and ELU-Net.
The number of GMAC operations increases as input resolution rises, affecting prediction scores.
Our architecture maintains higher prediction quality on lower input resolutions compared to other networks.
Цитаты
"Our architecture dramatically improves the prediction quality of the baseline while not adding any significant computations."
"Our experiments showed that our architecture can maintain higher prediction quality on lower input resolutions than other networks."
"Our architecture allows us to overcome limitations imposed by output resolution."