Topology-Aware Training and Spatial Diagnostics for Fiber Bundle Segmentation in Tracer Histology

SummaryAI generated

In the task of tracable histological fiber bundle segmentation in macaques, researchers compared the performance of the topological perception loss functions based on DINOv3 basic model features (Betti matching, Topograph) with conventional loss functions (BCE-Dice, clDice). The experimental results showed that BCE-Dice achieved the highest Dice coefficient; while Topograph achieved the lowest $eta_0$ error and fewer false positives while maintaining a similar Dice value. The study introduced the Excess32 spatial diagnostic index to quantify situations where predicted pixels exceeded the labeled boundaries. Validation revealed that although the combination of Betti-Topograph improved the recall rate (TPR) of sparse fiber bundles from 0.818 to 0.933, it led to an increase in false discovery rate (FDR) and Excess32 and area ratio. The study indicated that traditional detection rules cannot capture the problem of over-segmentation, proving that relying solely on detection metrics is insufficient to characterize segmentation quality.

Related eventsRELATED EVENTS

2026-09-09 05:29OpenAI claims huge maths breakthrough on a famed ‘Millennium Problem’
2026-09-09 06:11OpenAI claims to have solved the Navier-Stokes equation problem, but its originality is questioned by the academic community
2026-09-08 22:00Google DeepMind Maps 9 Billion Possible DNA Variants

Key entitiesKEY ENTITIES

BCE-DiceDINOv3TopographarXivclDice

Coverage · reports per dayLANGUAGE SPLIT

Entity relations

SignalsSIGNALS

Keyword heat

arXiv1
DINOv31
BCE-Dice1
clDice1
Topograph1

All reports (1)SOURCES

↗

A arXiv cs.CV en 2026-09-07 12:00

In the task of tracing tissue fiber bundles in macaques, researchers for the first time compared the performance of topological perception loss functions (Betti matching, Topograph) based on DINOv3’s basic model features with conventional loss functions (BCE-Dice, clDice). The results showed that BCE-Dice achieved the highest Dice coefficient, while Topograph achieved the lowest $\beta_0$ error and fewer false positives while maintaining a similar Dice value. The study pointed out that traditional detection rules cannot capture the problem of over-segmentation, and introduced the Excess32 spatial diagnostic index to quantify situations where predicted pixels exceed the labeled boundary. Verification showed that although the combination of Betti-Topograph increased the TPR of sparse fiber bundles from 0.818 to 0.933, it led to a deterioration in FDR, Excess32, and area ratio, proving that relying solely on detection metrics is insufficient to characterize segmentation quality.