Oral Leukoplakia (OLK) Segmentation Based On Weakly Supervised Learning

Oral leukoplakia(OLK)is a kind of precancerous lesion.Because of its visual similarity to healthy tissues in the oral cavity,it is difficult to be distinguished accurately.The current diagnosis of OLK mainly depend on the knowledge of the experienced doctors.However,this way is inefficient and the greatly affected by subjectivity.In this article,we will achieve automatic segmentation for oral leukoplakia,which can provide auxiliary guidance for doctors' diagnosis to facilitate the treatment and help patients to do self-diagnosis to avoid worse conditions.Additionally,the job for annotating the oral image dataset wants those experienced doctors,it is so costly that impacts the progress of the research.And the size of oral leukoplakia dataset is relatively small,the current weak supervised segmentation methods are not available.In this paper,we will focus on the shortcomings of the existing research methods and achieve excellent weakly supervised OLK segmentation performance.This article will start from the following aspects:1.Basic Segmentation method design based on fully supervised learning.In this part,we aim to address the problem about erroneous or missing segmentation with the traditional Mask R-CNN.After analyzing we think it is due to the weak feature extraction ability of the traditional Mask R-CNN,which makes model learn the invalid and wrong knowledge from the oral medical image.Recently,researchers design attention mechanism to help the network enhance the extraction of effective information and suppress the invalid.Therefore,we propose a Mask R-CNN oral leukoplakia segmentation method with attention mechanism.The spatial attention module from the convolutional block attention module is added into Mask R-CNN,so that the network can pay different attention to different positions on the feature map space.The impact of Non-OLK areas would be reduced and OLK areas would be augmented during network training process.In addition,due to the lack of data,the pre-trained model trained with COCO datasets is used,the trick " freezing shallow layers' parameters and updating deep layers' parameters" is adopted.Experiments show that our method achieves better segmentation performance than traditional Mask R-CNN,effectively solving the problem about erroneous segmentation and missing segmentation and generating the best fully supervised segmentation model for weakly supervised segmentation research to compare.2.Weakly supervised segmentation method design.To reduce high labor and time cost of oral leukoplakia labeling,in this paper we propose a weakly supervised segmentation method based on box annotation.However,there are some obstacles to clear,that are and poor localization ability of pixel-level coarse labels and low processing efficiency for it.They greatly affect the training and segmentation performance of the box-level weakly supervised segmentation model.To overcome the weakness,we will stick to work in the following two aspects.We propose an automatic pixel-level coarse labels generation method,which achieves efficient pixel-level coarse labels generation.We propose a collaborative optimization mechanism containing localization and segmentation based on Mask R-CNN with attention mechanism,which consist of the segmentation network training process based on box annotation and the process of pixel-level labels iteration.Through these methods,the localization ability of box annotation is sufficiently used and the accurate segmentation for OLK areas is effectively achieved.As experimental results show,compared with the model training process based on pixel-level coarse labels,the performance of the model training process based on box annotation is improved by 35.23%;after labels iteration,the segmentation performance of the weakly supervised segmentation model reaches 84.06% of the fully supervised segmentation model,when the labeling workload is only 14.13% of best fully supervised segmentation model.