Anchor-Free Feature Aggregation Network for Instrument Detection in Endoscopic Surgery

Abstract

Endoscopic endonasal approach has been widely used for removing various sellae tumors including pituitary adenomas, meningiomas, etc. While, performing these surgeries in such a narrow space with different instruments remains a challenge for surgeons, due to the limited field of view, varying illumination, and occlusion of instruments during the operation. Thus, a proper surgical instrument detection method that can provide classification and location information of the operated surgical instrument is critical for surgeons to understand the surgical scenarios and enhance the safety of the clinical operation. To this end, we propose an anchor-free feature aggregation network (AFA-Net) to improve the detection precision of surgical instruments from the endoscopic operation view field. The proposed method utilizes the improved feature pyramid network (FPN) with the depthwise separable convolution and a weighted feature aggregation module to enhance the feature information of the operated surgical instruments. Based on the anchor-free method, a weighted heatmap aggregation module is used to detect surgical instruments. Experimental studies on a public dataset Cholec80 and an intraoperative dataset from a local hospital are conducted, and the detection performance is assessed by the mean precision (AP) and average recall (AR). From both datasets and comparisons, the proposed method achieves 74.1% AP, 67.0% AR and 73.6% AP, 66.7% AR, respectively, which show significant advantages over five mainstream methods in terms of detection performance.

Keywords

• Endoscopic surgery
• surgical instrument detection
• convolutional neural network
• feature pyramid network
• anchor-free detection