Пошук академічних текстів

Інформація × Реєстраційний номер 2124U004688, Матеріали видань та локальних репозитаріїв Категорія Препринт Назва роботи Dense Vascular Segmentation in 3D scans of the human kidney with sparse annotation Автор Peleshko Severyn Дата публікації 01-01-2024 Постачальник інформації Український католицький університет Першоджерело https://hdl.handle.net/20.500.14570/4839 Видання Опис The main goal of this study is to investigate how various loss functions affect the problem of sparse segmentation, based on the task of kidney vasculature segmenta- tion. To achieve this we will conduct a comprehensive performance analysis of dif- ferent losses, examining both their weak and strong sides. The evaluated loss func- tions include Binary Cross-Entropy, its modification incorporating the Dice similar- ity coefficient(BCEDice), Focal loss, and Boundary Difference over the Union(BDoU) loss. Furthermore, this study will propose a novel loss criterion, which is a modified version of the standard BDoU loss(BDoUV2), incorporating an additional penalty for False Positive(FP) instances. Experiments showed that loss functions exhibit different performances while en- countering sparse segmentation issue, in particular, the poorest performance was achieved by the Focal loss, while the BDoUV2 loss along with the BCEDice loss showed the most prominent results. One common issue across these losses, to vary- ing degrees, is that utilizing each of them results in a huge amount of FN errors. Moreover, we have explored, that all loss functions, except BCEDice loss, are highly affected by the value of the classification threshold, with the Focal loss being the most sensitive to it. The insights regarding the performance of different loss functions under the task of sparse segmentation will be valuable for other scientists while developing an ac- curate and robust method for the task of medical sparse segmentation, where preci- sion is critical for effective diagnosis and treatment planning. All code used for this work can be found on the GitHub. Додано в НРАТ 2025-05-09 Закрити