Research On Segmentation Of Melanoma Image Based On Deep Learning

Skin cancer stands as the most prevalent form of malignancy,with melanoma constituting 75% of skin cancer-related mortalities,emerging as the most lethal subtype.Early-stage melanomas,however,can be effectively treated via minimally invasive surgical procedures.Consequently,the segmentation of dermatoscopic images holds significant clinical value by assisting dermatologists in promptly detecting melanocytic lesions,thereby enabling timely therapeutic interventions.Current methods for medical image segmentation often rely on high-quality datasets for research purposes.Although certain strategies consider clinical applicability,their focus remains largely on intrinsic dataset characteristics,disregarding the integration of external prior knowledge.This limitation results in inadequate resolutions to challenges posed by limited data availability,sparse annotated samples,and class imbalance,consequently leading to suboptimal performance.Therefore,this paper explores melanoma image segmentation research from the following aspects:(1)A novel weakly supervised edge-refinement technique is introduced for melanoma image segmentation,executed in three distinct stages.The first stage involves melanoma dataset classification,followed by weakly supervised methods to extract and process multi-tiered class activation maps.Subsequently,a UNet network architecture is constructed to segment malignant samples within the melanoma data set during the second stage.The third stage encompasses the utilization of augmented and overlaid class activation map edges to refine the segmentation outcomes.Experimental outcomes on the ISIC melanoma dataset demonstrate that the three-stage approach yields finer segmentation edges.Weakly supervised methods facilitate effective model training using sparsely annotated or incomplete datasets,thereby substantially lowering annotation expenses and addressing the issues arising from a lack of annotated samples and class imbalance.(2)The weakly supervised image classifier in this stage still does not fully address the aforementioned problems.A testing-time augmentation method suitable for skin lesions is proposed as a complement to the weakly supervised edge refinement method for melanoma image segmentation.This method enhances the classifier,improves model performance,and further addresses the issues of limited annotated samples and class imbalance.(3)Even with the use of weakly supervised methods to leverage benign image samples,the model still exhibits a dependency on a large number of malignant samples due to the strong privacy of medical images,making it difficult to obtain more data.A melanoma image segmentation method that integrates anatomical prior knowledge is proposed to further address the problem of insufficient data.By analyzing the anatomy of melanoma images,it is observed that the images with lesions present a star-like pattern.This medical prior knowledge is encoded into the UNet model’s loss function,and after segmenting various cases in a segmented form,the model is trained.Experimental results demonstrate that this model still performs well on datasets with insufficient data and holds practical value.In conclusion,the segmentation methods proposed in this paper effectively address the issues of insufficient data,limited annotated samples,and class imbalance,improving model performance and providing an effective solution for melanoma image segmentation.