Research And Application On Multi-expert Annotated Medical Image Segmentation Based On Deep Learning

In recent years,deep learning has swept the world and has been widely applied in various real-world tasks.In supervised learning tasks,deep neural network models are typically iteratively optimized with manually annotated training targets by domain experts,in order to achieve prediction capabilities close to those of experts.Thus,the availability of deep learning methods in the supervised learning paradigm depends on the uniqueness and authenticity of manual annotations.However,in the medical field,differences in experience and expertise among doctors often lead to frequent disagreements,resulting in inter-observer variability,then the "one-to-one" supervised paradigm is broken,making it difficult for deep network models to learn the correct patterns.In addition,this phenomenon also exists in medical crowdsourcing,where medical images are manually annotated by multiple amateur enthusiasts or professional annotators in a collaborative manner,supporting deep learning with big data.Especially in pixel-level medical image segmentation tasks,inter-observer variability is more significant due to the requirements for meticulous annotation.When facing the segmentation learning problems with multi-expert annotations,classical deep learning algorithms are unable to directly learn from the multi-annotation,and cannot cope with real-world scenarios where multiple doctors are involved in segmentation annotations or crowdsourcing segmentation learning.Existing label fusion methods and models for learning from multiple expert annotations also have limitations and room for improvement.Therefore,multi-expert annotated medical image segmentation methods have significant research significance.Given the context described above,this thesis conducts research on multi-expert annotation medical image segmentation methods and applications and proposes robust end-to-end models based on convolutional neural network.The main contributions are as follows:(1)This thesis designs a crowdsourcing segmentation process based on fundus image data and constructs a crowdsourcing segmentation dataset for the optic disc and cup,which is a dataset where multiple amateur annotators independently annotate the multi-color fundus image set after being trained by ophthalmologists.This work uses the collective wisdom of non-medical experts to provide data support for deep learning training processes,promoting cost reduction and efficiency improvement in constructing medical image segmentation datasets.(2)This thesis proposes a multi-expert annotated segmentation deep model based on deep learning algorithms and trace regularization theory,which models annotator bias and reliability through a trace-regularized network constrained by a piece-wise loss function.The regularization network is jointly optimized with the segmentation network to enable the model to automatically learn useful information from multiple annotations in the training process.The model performs exceptionally well in the proposed crowdsourced dataset for optic disc and cup segmentation of color fundus images,and shows potential to perform on par with expert models.(3)This thesis combines the soft label learning strategy and spatial smoothing method with weakly supervised learning strategy,and proposes a local consistency regularized soft label learning method that quantifies the uncertainty reflected in multi-annotation as the inter-class variance and uses it as prior information to guide the learning of local variability.Compared with traditional label fusion algorithms and other deep learning methods,the proposed method performs exceptionally well in realistic publicly available datasets with multi-expert annotations for the segmentation of optic disc and cup,prostate,and kidney.In summary,this thesis proposes two multi-expert annotated medical image segmentation methods based on deep learning in the context of differential annotation problems for multi-doctors and crowdsourcing learning.Ultimately,they achieved segmentation accuracy of93.75%/93.92%,95.08%/94.51%,70.39%/72.25%,and 87.07%/87.67% in the RIGA,CSD,QUBIQ-Kidney,and QUBIQ-Prostate datasets respectively,surpassing the commonly used and optimal methods in most evaluation indicators.This thesis provides a new solution with performance advantages for multi-expert annotated medical image segmentation problem.