Weighted Circle Fusion: Ensembling Circle Representation from Different Object Detection Results

TL;DR

The paper tackles the lack of circle-specific ensembling for spherical object detection in medical imaging and introduces Weighted Circle Fusion (WCF), which pairwise merges circle detections from five CircleNet-based models using circle IOU and confidence-weighted averaging. Key computations include $x_{fuse} = (x_1 s_1 + x_2 s_2)/(s_1+s_2)$, $y_{fuse} = (y_1 s_1 + y_2 s_2)/(s_1+s_2)$, and $r_{fuse} = (r_1 s_1 + r_2 s_2)/(s_1+s_2)$, with the final fused score $S = \frac{\sum S_m}{M}$ and a count $C$ of merged detections; detections are retained only if $C \ge T_{count}$ and $S \ge T_{score}$, where $T_{count}=2$ and $T_{score}=0.9$. Evaluations on renal and murine whole-slide images demonstrate >5% improvement in mean AP over individual models, enhanced rotation consistency, and substantial time savings when using a HITL annotation approach (≈68.6% faster). The method, along with its public implementation, offers a practical path to more reliable circle-based object detection in pathological image analysis. The results underscore WCF’s potential to reduce false positives and improve detection precision in circle representations.

Abstract

Recently, the use of circle representation has emerged as a method to improve the identification of spherical objects (such as glomeruli, cells, and nuclei) in medical imaging studies. In traditional bounding box-based object detection, combining results from multiple models improves accuracy, especially when real-time processing isn't crucial. Unfortunately, this widely adopted strategy is not readily available for combining circle representations. In this paper, we propose Weighted Circle Fusion (WCF), a simple approach for merging predictions from various circle detection models. Our method leverages confidence scores associated with each proposed bounding circle to generate averaged circles. We evaluate our method on a proprietary dataset for glomerular detection in whole slide imaging (WSI) and find a performance gain of 5% compared to existing ensemble methods. Additionally, we assess the efficiency of two annotation methods, fully manual annotation and a human-in-the-loop (HITL) approach, in labeling 200,000 glomeruli. The HITL approach, which integrates machine learning detection with human verification, demonstrated remarkable improvements in annotation efficiency. The Weighted Circle Fusion technique not only enhances object detection precision but also notably reduces false detections, presenting a promising direction for future research and application in pathological image analysis. The source code has been made publicly available at https://github.com/hrlblab/WeightedCircleFusion

Figures (3)

• Figure 1: Comparison of Box Fusion and Circle Fusion Methods for Object Detection. This figure delineates the differences between the ensemble results of box representation and circle representation. Box fusion alters the dimensions of the box, thereby changing its shape, while circle fusion only modifies the radius of the circle, preserving its shape. For the detection of medical ball-shaped objects, circle representation can achieve better performance.
• Figure 2: The workflow of the proposed Weighted Circle Fusion (WCF) method. This figure delineates the specific steps involved in our method. The core of the method lies in counting the number of fused circles and calculating their average score, which is then used to eliminate potential erroneous detections.
• Figure 3: Result Visualization. This figure presents the detection outcomes of glomeruli on WSIs using our method. The yellow arrows highlight false negatives identified by other models or methods, while the blue arrows indicate false positives. It is evident that traditional fusion methods such as NMS and soft-NMS tend to merge more erroneous predictions. In contrast, the WCF method achieves superior fusion results, with fewer incorrect predictions and the inclusion of detections that individual models failed to identify, demonstrating its effectiveness in enhancing detection accuracy.