Abstract
Positive surgical margin assessment in computational pathology is a low signal-to-noise task because benign inflammation, atrophic glands, and other histological mimics may resemble high-grade carcinoma in restricted fields of view. We propose TD-VAN, a top-down visual attention framework that treats coherent global tissue architecture as a beneficial signal rather than nuisance variation. An Overview-Net extracts low-power architectural and topographical priors, including glandular crowding, infiltrative arrangement, stromal organization, inflammatory aggregation, spatial continuity of suspicious glands, and their relationship to the inked margin. These priors dynamically guide a Focus-Net for context-aware local recognition, helping suppress false positives caused by deceptive mimic noise. Multi-scale sampling integrates architectural and cytological evidence, while automated Topographical Zonal Profiling converts predictions into spatially resolved metrics, including tumor burden, minimum tumor-to-margin distance, and Gleason-grade composition. On a temporally held-out cohort from 2022-2024, TD-VAN achieved an accuracy of 0.971 and a Macro F1-score of 0.974. These results indicate that global context can stabilize surgical-margin assessment and provide interpretable evidence for postoperative risk stratification.
| Original language | English |
|---|---|
| Article number | 110886 |
| Journal | Signal Processing |
| Volume | 251 |
| DOIs | |
| State | Published - Feb 2027 |
UN SDGs
This output contributes to the following UN Sustainable Development Goals (SDGs)
-
SDG 3 Good Health and Well-being
Keywords
- Computational pathology
- Deep learning
- Positive surgical margins
- Prostate cancer
- Risk stratification
- Visual attention network
Fingerprint
Dive into the research topics of 'Exploiting global context as beneficial signal instead of noise: A top-down framework for robust prostate margin processing'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver