LIVER DETECTION ALGORITHM BASED ON LOCAL INFORMATION FUSION

Abstract

The liver is one of the vital organs of the human body, and its location detection is of great significance for computer-aided diagnosis. There are two problems in applying the existing algorithms based on convolution neural network directly to liver detection. One is that pooling operation in the convolutional layer, characteristic of the existing algorithms, will result in local information loss, and the other is that direct calculation of area-based pre-defined anchor boxes will cause incomplete alignment of the generated anchor boxes with overall data distribution. As a solution, this paper suggests a liver detection algorithm based on local information fusion. First, area calculations are complemented with the target aspect ratio as a constraint term to generate a predefined anchor box more in line with actual data distribution. Second, the local feature fusion (LFF) structure is proposed to bridge local information loss caused by pooling operation. As the final step, LFF is used to optimize the neural network analyzed in YOLOv3 for liver detection. The experimental results show that the optimized algorithm achieves an average intersection over union (IoU) in liver detection three percentage points higher than the YOLOv3 algorithm. The optimized algorithm proves more accurate in portraying local details. In the object detection of the public data set, Average Precision for medium objects (APm) and Average Precision for large objects (APl) are 2.8% and 1.7% higher than their counterparts derived from the YOLOv3 algorithm, respectively.