Research On Skin Cancer Assisted Diagnosis Method Based On FixCaps

Diagnosing skin lesions accurately is a crucial step in screening for skin cancer.While deep learning-based methods have made significant progress in this area,their stability and generalization remain challenging.Existing methods often struggle to handle the large variations within a single class of skin lesions and the high similarities between different classes.Consequently,they may exhibit redundant features and suboptimal performance in clinical settings.In order to address these issues,the main research content of this thesis is as follows.1)The proposed method,called FixCaps,aims to address the challenges posed by intra-class variation and inter-class similarity in skin lesions for skin cancer auxiliary diagnosis.It is based on attention mechanisms and an improved capsule network.To enhance feature extraction,the algorithm utilizes a large kernel convolution with a kernel size of up to 31x31 in the initial layer.This larger receptive field surpasses the 9x9 convolution kernel used in Caps Nets,enabling the network to capture more comprehensive feature information.Furthermore,a convolutional attention module is introduced to mitigate spatial information loss caused by pooling operations.Additionally,group convolution is employed to prevent model underfitting.Experimental results demonstrate that our algorithm achieves a skin lesion diagnosis accuracy of96.49% on the HAM10000 dataset,surpassing IRv2-SA by 3%,while reducing computation by 73.53%.2)To further enhance the stability and robustness of FixCaps,an improved algorithm called FixCaps V2 is proposed.The algorithm incorporates asymmetric convolution and global max pooling to enhance the capsule architecture.Additionally,it transforms the output form of the model from a vector to a matrix and utilizes kernel norm to measure the low-rank nature of the output matrix.Furthermore,to meet a broader spectrum of pathological evaluation requirements,a novel margin loss function is proposed,and improvements are made to the squash function and dynamic routing,thereby expanding the model’s capability.The experiments have shown that FixCaps V2 not only outperforms FixCaps in the diagnosis of skin lesions,with an accuracy of 99.37% on the HAM10000 dataset,but also balances stability and robustness.It has been successfully applied to the COVID-19 and NCT-CRC-HE datasets with accuracies of 99.72% and 99.91%,respectively.The algorithm proposed in this thesis not only achieves accurate diagnosis of skin lesions,but also has excellent generalization and stability,providing a more reliable and efficient solution for diagnosis and treatment in the medical field.Experimental results demonstrate that the algorithm can assist medical personnel in diagnosing multiple diseases efficiently,achieving significant results on the three datasets of HAM10000,COVID-19,and NCT-CRC-HE.The application of this algorithm will improve the early screening rate of diseases such as skin cancer and colorectal cancer,further increasing the five-year survival rate of patients,and positively impacting the quality of life and health status of patients.Therefore,the proposed algorithm proposed in this thesis has important application value and prospects.