A Computer Vision–Based System for Automated Counting of Surgical Supplies
This case study proposes a multi-stage deep learning-based system for an automated inventory analysis of stent boxes cabinets in angiography rooms. The proposed pipeline integrates cabinet image segmentation, vendor classification, detection of region-of-interest (ROI) with numerical features, and character recognition enabling extraction of stent boxes’ attribute triples (vendor, diameter, length). Experimental results show that the system achieves high performance across individual stages: mAP@0.5 of 0.995 in box segmentation task, top-1 accuracy of 99.3% in the stent vendor classification and mAP@0.5 of 0.991 in ROI detection. The overall F1 score at the system level for stent box attribute triples is 0.811. The error analysis indicates that system performance is strongly influenced by camera-to-cabinet distance determining the scale of segmented ROI, and can be improved to F1=0.984 by partial standardization of image acquisition step.