Small Object Detection Method For CT Images Of Liver Tumors Based On Adaptive Training Sample Selection

Small object detection for CT images of liver tumors is an important part of smart medical construction.However,due to the characteristics of tumors such as large scales,many small objects,sparse distribution,blurred object edges,and complex backgrounds,the detection of small tumor in liver CT images is still a very challenging task.In this paper,we propose a small object detection model called IPSC-ATSS for liver tumor object detection.This model using the adaptive training sample selection strategy to promote detection effectiveness.The model focuses on the bounding box quality evaluation,loss function design and backbone network framework.These improvements avoid low-quality prediction bounding box,unbalanced sample and features at different scales that appear in the task of small object detection in liver tumor CT images,so as to improve the performance of small object detection in liver tumor CT images eventually.The main analysis content are as follows:(1)We design a data preprocessing procedure for CT images of liver tumors.Through resampling,filtering,gray-scale and equalization operations,the contrast of liver tumors is improved and the interference of background noise is suppressed.(2)We design an anchor frame for liver tumor object detection models,and constructs a single-stage liver tumor detection model framework called IPSC-ATSS.The model using image pyramid residual network(IP-ResNet)and self-calibration feature pyramid(SC-FPN)to deal with the uneven information of feature maps of different layers.Focal loss and GIoU loss are used to solve the problem of model training instability and sample imbalance.Finally,the output uses the bounding box quality evaluation branch,and uses the classification regression network based on ATSS to improve the quality of the predicted bounding box.(3)Ablation experiments confirmed that the IPSC-ATSS has better detection performance for small liver tumor objects.The AP evaluated on the LiTS dataset reached 44.9%,and the APs reached 31.5%,which is better than the existing algorithms,and the detection speed reached 19.7FPS,which has comprehensive advantages.