CAM Based Fine-grained Spatial Feature Supervision On Surgical-PPE: A New Dataset For Surgical PPE Kit Presence Detection

Abstract

In the wake of Covid pandemic, usage of surgical PPE kit by surgeons has become essential. Since reliable localization of human joints is necessary for automated understanding of surgeons activity, the first step is surgical PPE kit detection. While there exist reported works on industrial PPE kit detection, task of surgical PPE kit detection has hardly been explored. To facilitate this, we construct "Surgical-PPE" dataset with 1150 Non-PPE instances and 2656 surgeon wearing PPE kit instances. In this work, we also propose a two-stage transfer learning based end-to-end training methodology. Novelty lies in (a) novel "Surgical-PPE" dataset to detect if surgeon is wearing PPE kit or not, (b) proposed supervised contrastive combined loss function for stage-1 training, (c) proposed spatial context aware combined loss function for stage-2 training. We qualitatively illustrate the improvement of HiResCAM and XGrad-CAM explanations for the proposed methodology. We also qualitatively illustrate that feature embeddings of same class are pulled closer together compared to feature embeddings of different classes on the proposed multi-stage training methodology, using T-SNE plots. We benchmark the performance of popular existing network architectures along with the proposed methodology on "Surgical-PPE" dataset. Using proposed methodology, we achieve peak accuracy of 97.63%, precision of 97.66%, recall of 97.63%, F1-score of 97.64%, JI of 95.41% and FPR of 2.5%. We report improvement by 1.7% in terms of FPR and 2% in terms of JI compared to second best performing model (ResNext50(CE)). Owing to the proposed training methodology, an improvement of 2.62% in terms of FPR and 5% in terms of JI was observed.Clinical relevance- To understand the OT activity of the surgeon in third-person perspective, it is important to determine whether or not, the surgeon is wearing PPE kit. Hence surgical PPE kit presence detection becomes the first step towards automated surgical video analysis.

Ähnliche Arbeiten