The Improvement Of Cascade Mask R-CNN Model And Its Application Research In The Recognition Of Benign And Malignant Thyroid Nodules

Aiming at the problems that the existing methods of detecting and recognizing thyroid nodules in ultrasonic images cannot accurately locate and the object detection accuracy is not high,based on multi-task deep learning,we propose a model of automatic detecting and recognizing for thyroid nodules based on the improved Cascade Mask R-CNN.First of all,we improve the Cascade Mask R-CNN network from three aspects.The first is to improve the Cascade Mask R-CNN by designing a more effective detector to better classify the ROI area and correct the bounding boxes(bbox),which aims to solve the problem that when the model is directly applied to the detection and recognition of thyroid nodules in ultrasonic images,it cannot be accurately located and the object detection accuracy is not high.The second is using a more effective balanced L1 loss function to increase the gradient of the accurate sample,which aims to solve the problem of imbalance between classification and locating tasks in the objective function during training of thyroid ultrasound images.The third is using a more effective Soft-NMS method to set an attenuation function for the adjacent bounding boxes,which aims to solve the problem that a real object may be undetectable when it is within the preset overlap threshold.Secondly,for the problem that the number of samples is not enough to train a perfect model independently,we propose a method of transfer learning,the pre-processed samples are input to the pre-trained ResNet-101 for parameter fine-tuning,and construct an automatic detecting and benign and malignant recognizing model for thyroid nodules based on improved Cascade Mask R-CNN.The proposed model was trained and verified by using 1408 images collected from the hospital.Under the locating accuracy of the IOU threshold of 0.5,the mAP reached 87.1%,and the object detection accuracy reached 98.67%,the results show that the model is effective.Finally,an easy-to-operate detecting and recognizing system for benign and malignant thyroid nodules in ultrasonic images is developed based on the model,aiming to assist clinicians to locate and recognize the nodules in thyroid ultrasound image and reduce the labor intensity.