Abdominal Organ Detection And Recognition In Ultrasound Images Based On Improved YOLOX And Knowledge Distillation

Abdominal ultrasound is a very important basic content in ultrasound diagnosis.It can detect some common diseases in time and take necessary measures to deal with them,which plays an important role in protecting people’s health.Detecting and identifying individual tissues and organs on ultrasound images often requires a physician’s expertise to make a decision.Utilize the latest advanced image processing technology to analyze and process ultrasound images with high precision,so as to provide doctors with more accurate and reliable diagnostic results,and make the detection of abdominal organ regions more accurate and faster.The method of deep learning can quickly help inexperienced doctors quickly find abdominal organs on ultrasound images,and reduce the burden of experienced doctors observing images for a long time.In this paper,some explorations have been made in abdominal ultrasound image detection,the main work is as follows:(1)For the problem of insufficient samples in the abdominal ultrasound dataset.By using Label Me to calibrate the abdominal ultrasound segmentation data set(USsimulation&segmentation),382 pieces of target detection data are calibrated,which contain the location information and category information of the abdominal tissue area,so as to realize the labeling of the target detection task.The data set is extended to 1528 images using data augmentation methods.(2)Aiming at the problem that the effect of the detector is affected by the change of the organ with the breath,a new abdominal organ detection model based on YOLOX is proposed.YOLOX has excellent industrial application capabilities,its detection speed is fast,and its accuracy is extremely high.However,when encountering ultrasonic images of abdominal organs with noise and complex morphological changes,YOLOX’s performance cannot reach the ideal level.In this paper,we discuss how to combine a heuristic algorithm with local spatial variation-Deformable Convolutional Network(DCN)with YOLOX to obtain a new model DF-YOLOX,which has better adaptability and detection results.Experiments have proved that the application of the DF-YOLOX model in the field of abdominal organ detection has significantly improved its detection accuracy,far exceeding the traditional benchmark model.(3)In order to solve the problem of too many model parameters and too slow running speed in the abdominal organ detection task,a new YOLOX-tiny(KD)abdominal organ detection algorithm is proposed,which considers both local factors and overall factor.The algorithm uses knowledge distillation to compress the model,and solves the problem of unbalanced foreground and background targets through the local method of foreground and background separation,and the global method of GCBlock to solve the problem of disassociation of foreground and background context information.In order to improve the real-time performance,the YOLOX-tiny network is selected as the student network,and the abdominal organ detection network of Chapter 3 DF-YOLOX is used as the teacher network.The training of the YOLOXtiny network is guided by local and global knowledge distillation,and finally a new YOLOX-tiny(KD)abdominal organ detection algorithm is proposed.Its detection rate can reach 67 FPS,which greatly improves the real-time performance and performance.accuracy.After precise measurement,the accuracy rate of abdominal ultrasound detection has reached 92.51%,which meets the detection speed and applicability requirements of abdominal ultrasound detection in the medical diagnosis environment.