The Research On Automatic Segmentation Algorithm Of Medical Images Based On Deep Learning

Medical imaging is a very important tool to help clinicians understand the patient’s condition and assist doctors in formulating treatment plans.While manual evaluation of medical images improves the accuracy of diagnosis,the process is often time-consuming and complex.Subjective interpretations of pathology may also vary with the level of knowledge of multiple observ ers.Medical image segmentation technology has gradually become a popular auxiliary diagnosis method due to the improvement of analysis efficiency and objectivity.It plays a vital role in the early and accurate diagnosis of many diseases.However,the texture,location,outline,color,and size of lesions are also difficult to distinguish due to the low contrast between the lesion area and the normal area.Automatic segmentation of medical images remains a challenging task.Therefore,in view of the above-mentioned problems,developing an accurate and efficient automatic medical image segmentation algorithm to assist clinical diagnosis is a subject with important research and application value,which is also the research subject of this paper.These research works can be summarized as:(1)A cross-level information fusion medical image segmentation network framework was proposed to improve the multi-level context information fusion ability and feature extraction ability of the medical image segmentation model.This is an end-to-end deep learning framework for medical image segmentation.The proposed framework includes an attention-based hybrid residual convolution module,a multi-scale feature memory module,and a multi-receptive field fusion module,which can fuse features at different levels and effectively extract contextual information.In this paper,the accuracy of the proposed cross-level information fusion medical image segmentation network model is verified in the tasks of skin melanoma lesion segmentation and retinal blood vessel segmentation.(2)A new feature compression pyramid network was proposed to solve the stability problem of attention weights and the underutilized complementarity between attention mechanism and residual blocks,and multi-level features are rarely seamlessly integration into different attention mechanisms,which may lead to the challenge of redundant use of low-level features.We propose a new strategy in the encoding stage,including an embedded feature integration module,an extended spatial mapping and channel attention module,and a branch-level fusion module.These modules effectively extract spatial information,efficiently capture channel and spatial correlations between features,and integrate multi-scale information from different feature branches,thereby improving the performance of lesion segmentation.In the decoding stage,we perform multiple multi-scale fusions compared to the singlestep fusion of the previous Deeplabv3+ network.To solve the information redundancy problem brought by fusion,we use extended spatial mapping and channel attention module for lightweight feature extraction.This paper validates the accuracy of the proposed new feature compression pyramid network model on three different datasets for the skin melanoma lesion segmentation task.(3)A multi-task learning-based feature compression pyramid network framework was proposed to further explore the potential benefits of the interaction between medical image segmentation and classification tasks for medical image segmentation tasks.This framework introduces a classification branch and an interaction branch in the segmentation task,exploring the interaction between medical image segmentation and classification.A new hybrid loss function is proposed from a game theory perspective.On this basis,we can make the segmentation,classification and interaction branches learn and teach each other synergistically throughout the training process,thus making full use of joint information and improving generalization performance.In this paper,the effectiveness of the proposed multi-task learning-based feature compression pyramid network framework is validated in skin melanoma lesion segmentation,polyp segmentation,ultrasound-based breast lesion location segmentation,and optic disc segmentation tasks.We conducted a large number of experimental analysis on the above methods on multiple public medical imaging datasets,which strongly proves its excellent performance on medical imaging tasks.Among them,the designed feature compression pyramid network framework based on multi-task learning The performance of the proposed model on these datasets outperforms the performance of other current deep models.We then visually analyze the stability of the weights in the attention mechanism and visualize the feature maps of the network to understand the performance of the network.Finally,the proposed method is summarized and the direction of future medical image analysis research is discussed.