Deep Learning-based Image Segmentation Methods In The Treatment Of Benign And Malignant Uterine Tumor Diseases

The uterus is a pear-shaped hollow organ in the female pelvis,located between the bladder and the rectum.Uterine tumors may occur in the body,isthmus,and cervix and may be benign or malignant(cancerous).Among these,benign tumors,represented by fibroids,and malignant tumors,represented by cervical cancer,are currently becoming a serious health risk for women.With the development of medical technology,there have been more advanced surgical treatments for benign and malignant tumors of the uterus.Computer-aided methods to improve surgery’s accuracy,efficiency,and safety are essential to women’s health.In the treatment of benign and malignant tumors of the uterus,accurate annotation of the lesions in the uterine region and the surrounding crisis organs is an essential part of the diagnosis and treatment planning: 1)In treating uterine fibroids,lesion annotation helps the surgeon determine the fibroid’s size,shape,location,and,thus,the type of fibroid.2)When treating adaptive radiotherapy(ART)procedures for cervical cancer,the doctor can develop the radiotherapy process and the prescribed dose based on the results of the delineation.3)In highintensity focused ultrasound surgery(HIFU)for uterine fibroids,the target area in the preoperative images is mapped to the intraoperative images to guide the surgery.However,annotating target areas and organs at risk in medical images are time-consuming and laborintensive,and factors such as image noise make accurate annotation difficult.Therefore,exploring automatic and accurate annotating methods for uterine images has significant clinical value for treating benign and malignant uterine diseases.To this end,this paper investigates the multimodal image segmentation algorithms and automated surgical techniques involved in HIFU and ART to treat uterine fibroids and cervical cancer,respectively.This paper aims to improve the automation and precision of these two treatments in clinical practice.The main work and contributions are as follows:1.Multi-class segmentation of uterine regions from MR images using global convolutional networks for HIFU surgery planningTo address the problem that existing state-of-the-art(SOTA)deep learning segmentation methods are not effective enough for complex multi-level feature extraction,we propose a novel convolutional neural network called HIFUNet to segment the uterus,uterine fibroids,and spine.The network is an end-to-end encoder-decoder architecture with a global convolutional network(GCN)module to expand the valid receptive field and extract multi-scale contextual information.In addition,combining GCN with our proposed deep multiple atrous convolution(DMAC)module can further extract contextual semantic information and denser feature maps.Our approach is compared to both conventional and other deep learning methods and the experimental results conducted on a large dataset show its effectiveness.2.Semi-supervised uterine MR image segmentation method based on pseudo-label Refinement for HIFU procedure planningThe fully-supervised semantic segmentation approach is effective but requires a high amount of annotation on the training data,so we propose a semi-supervised deep learning approach to segment MRI images.This method aims to refine the pseudo-label generated in the semi-supervised method named: pseudo-label refinement network(PLRNet).Inspired by the fully supervised uterine segmentation method HIFUNet,feature extraction can be improved by expanding the valid reception field in segmenting uterine fibroids with different sizes and shapes.Therefore,in semi-supervised feature extraction,we consider a network with large convolutional kernels to extract contextual features for each target class in MR images.In addition,the feature dilution problem caused by the typical pooling operation in deep learning is improved,and wavelet pooling is utilized to suppress the image noise.Unlike the traditional semi-supervised approach of setting fixed thresholds,we use an adaptive method based on confidence thresholds for the first time in semi-supervised segmentation to improve the quality of the pseudo-label.In addition,inspired by "Mixup" method,we extend Mixup operations to each hidden layer of the fine segmentation network,which helps data augmentation and avoid the overfitting phenomenon that tends to occur in semi-supervised learning,improving the generalization and robustness of the model.3.Automatic segmentation for plan-of-the-day selection in CBCT-guided adaptive radiation therapy of cervical cancerPlan-of-the-day(PoD)-based ART is based on a library of treatment plans.At each treatment fraction,the PoD is selected based on daily images.However,this strategy is limited by the optimal PoD selection due to visual uncertainties.This work proposes a workflow to automatically and quantitatively determine the PoD of ART for cervical cancer based on daily CBCT images.The quantification is performed by segmenting the main structures of interest(Clinical target volume(CTV),rectum,bladder,and bowel bag)in CBCT images using a deep learning model.Then,the PoD is selected from the treatment plan library according to the geometrical coverage of the CTV.The resulting PoD is compared to the one obtained considering reference CBCT delineations for the evaluation.