Multi-Tasking DSSD Architecture for Laparoscopic Cholecystectomy Surgical Assistance Systems

Abstract

In this paper, we propose a novel DSSD based encoder-decoder multi-tasking architecture for the simultaneous tasks of (i) surgical tool presence detection, (ii) surgical tool localization and (iii) surgical phase classification - all on laparoscopic cholecystectomy surgical videos for the purpose of visual surgical assistance. Novelty of the study lies in addressing all the three tasks simultaneously with a single network architecture. Peak performance was achieved on m2cai16-tool-locations dataset at 97.51% mAP for the task of surgical tool presence detection, 91.9% mAP for the task of surgical tool localization (20% higher than SOTA), 97.77% accuracy for the task of surgical phase classification. This multi-tasking approach reduces the demand over training images needing only 2025 training images as against 2.3M images required otherwise. Besides, the approach needs only less than 30% of the model parameters than those that perform each of these tasks separately.