Research On Medical Image Segmentation Algorithm Based On Multi Model Adaptive Fusion

With the continuous development of medical imaging technology and computer vision technology,medical image segmentation plays a key role in lesion assessment.In clinical medicine,the results of medical image segmentation directly affect the diagnosis of patients conditions and subsequent treatment plans.However,medical image morphology is complex and variable,and manual interpretation requires a lot of time,and different doctors have different subjective judgments,making the medical image labeling results lack objectivity.Therefore,there is an urgent need for algorithms that can automatically segment medical image structures to assist doctors in accurately diagnosing patients conditions.This paper takes retinal vessel images and skin lesion images as research samples,perform multi-model adaptive fusion of deep learning methods,and proposes three medical image segmentation algorithms: ghost convolutional adaptive retinal vessel segmentation algorithm,skin lesion segmentation algorithm based on high-resolution composite network,and adaptive multi-scale Transformer skin lesion segmentation algorithm.The main research content and innovative work are as follows.(1)A ghost convolutional adaptive retinal vessel segmentation algorithm is proposed to address the challenges of blurred main vessel contours,microvessel ruptures,and missegmentation of the optic disc boundary in retinal vessel segmentation.Firstly,the ghost module is used to replace ordinary convolutional layers in the neural network,which generates rich feature maps in a step-by-step output manner to expand the information of the target area.Secondly,the encoder generated feature maps are adaptively fused and input to the decoder for classification prediction.The adaptive fusion module uses convolutional kernels with different receptive fields to obtain multi-scale features,allowing for better transfer of retinal microvascular features.Thirdly,a dual-path attention guidance structure is constructed to fuse the low-level feature maps in the encoder with the high-level feature maps in the decoder to improve retinal vessel segmentation accuracy by accurately locating vessel pixels and addressing image texture loss.Experiments on the DRIVE and STARE datasets demonstrate that the ghost convolutional adaptive network has good segmentation performance for retinal vessel images and provides new ideas for the diagnosis of ophthalmic diseases.(2)A skin lesion segmentation algorithm based on high-resolution composite network is designed to address challenges such as foreign object occlusion,feature loss,and mis-segmentation of lesion areas in skin lesion image segmentation.The algorithm uses a high-resolution network and a multi-scale dense module to construct the encoding part.The structure of the high-resolution network connected in parallel can repeatedly fuse feature maps of different resolutions to ensure the global transmission of high-resolution features.The multi-scale dense module can capture dense and discretely distributed targets,reducing the loss of lesion features.Meanwhile,the decoding part is constructed using a reverse high-resolution network and a dual residual module.The dual residual module can understand deep semantic information when reconstructing the spatial dimension information of the image,accurately locating the lesion area.Additionally,the algorithm applies morphological operations to refine skin images and reduce the impact of foreign object occlusion on segmentation performance.Finally,experiments were conducted on the ISIC2016,ISIC2017,and ISIC2018 datasets,and the results demonstrate that the high-resolution composite network has a good segmentation effect on skin lesion images,providing a new window for the diagnosis of skin diseases.(3)Considering the fusion of multiple fields in computer vision,to deal with complex morphology and structure of skin lesions,a skin lesion segmentation algorithm with an adaptive multiscale Transformer encoder-decoder network is constructed by integrating the Transformer structure from natural language processing into medical image segmentation.Firstly,a hierarchical encoder is built using Transformer Blocks,which analyzes the skin lesion area from the perspective of global feature changes at multiple scales.Secondly,a fusion decoder is constructed using a multiscale fusion module,channel attention module,and joint layer,which complements the shallow and deep network information in the hierarchical encoder and enhances the dependence between semantic information.The channel attention module processes the feature maps output from the multiscale fusion module,where the more informative feature maps have a higher weight,retaining the key content in the image.Then,an extension module with multi-level output is used to integrate features from different stages and restore the image scale to match actual requirements.Finally,experiments were conducted on the ISIC2018 and PH2 datasets,and the results proved that the adaptive multiscale Transformer encoder-decoder network can effectively segment skin lesion images,providing new inspiration for the fusion of multiple fields in computer vision.