SurgStreamer times alerts for events in surgical video
Esta página aún no tiene una traducción verificada. Se muestra la versión en inglés. Leer la versión en inglés
En palabras sencillas
This paper introduces SurgStreamer, a system that watches surgical video for requested events. It turns the video into records of which tools, actions and procedure phases are present, then checks those records to decide when to alert or answer questions about what happened earlier.
On 12 sealed procedures, its hybrid system matched alerts to annotated events better than exact rules did. The test used prewritten request templates and reference conditions, so it did not show how surgeons would use or rate the alerts.
La investigación
A surgical team may need an alert when a requested event is confirmed, or an answer about what has already happened. For such a monitor to be useful, its timing matters: it should alert once for each requested occurrence and remain silent when the event did not occur. Can a system meet those requirements from surgical video?
SurgStreamer separates surgical video perception from event reasoning, then checks structured state to time requested alerts and answer questions about earlier events.
The main test asks how accurately SurgStreamer times alerts compared with exact rules. Macro event-F1 matches alerts to annotated occurrences and averages scores across proactive query families.
On 12 sealed procedures, the hybrid scored 0.598 and exact rules scored 0.570 on the 1,024-second state. The hybrid selected a state and controller for each query family using training-fold results.
A learned controller on the 1,024-second state scored 0.607 and was not significantly different from the hybrid. This supports an advantage over the exact-rule comparison under the tested conditions; it does not establish the hybrid as the best variant.
The main evaluation used template-generated requests and gold monitoring conditions, so it tested perception and event reasoning rather than request compilation. No surgeons wrote requests or rated alerts, and there was no prospective evaluation.
Análisis
Detalles
- Autores del paper original
- Bowen Xiang, Jon S. Heiselman, Michael I. Miga
- Enviado a arXiv
- Licencia
- arXiv non-exclusive