优化ViT用于黑色素瘤分类：特征筛选与InfoNCE损失的结合

打开文本图片集

关键词：图像分类；特征筛选；InfoNCE损失函数；ViT模型中图分类号：TP394.1文献标识码：Adoi：10. 37188/OPE.20253316.2649 CSTR：32169.14.OPE.20253316.2649

Optimizing vision transformer for melanoma classification ： integration of feature selection and InfoNCE loss

HUANG Jinjie*，MAYuanxue

（School ofAutomation，Harbin University ofScience and Technology，Harbin 15O08O，China） * Corresponding author， E -mail： jjhuang@hrbust. edu. cn

Abstract：To address the issues of feature redundancy and insuficient generalization in Vision Transformer（ViT） models for melanoma image classification，we proposed an enhanced model that integrated dynamic feature selection and contrastive learming. This approach aimed to improve classification accuracy and clinical diagnostic eficiency. Specifically，a dynamic feature selection module was introduced，which adaptively enhanced key features while suppressng redundant information using a learnable weight matrix. Additionally，the InfoNCE contrastive loss function was incorporated into a multi-objective optimization framework alongside cross-entropy loss，thereby improving the discrimination of inter-class features. Furthermore，a feature importance guidance mechanism was embedded within the multi-head self-attention mechanism to achieve collaborative modeling of local details and global semantics.Experimental results on the ISIC2O18 and ISIC2Ol9 datasets demonstrate that the improved model achieves clasification accuracies of 83.27% and 80.17% ，respectively，surpassing the baseline ViT by 1.83% and 0.49% . Ablation studies confirm that the dynamic feature selection module reduces computational redundancy by 18.7% ， whilecontrastive learning increases intra-class feature similarity by 23.6% .The proposed method significantly enhances the recognition capabilities of the ViT model for melanoma，ofering superior clasification accuracy and robustnesscompared to mainstream models. It provides a high-precision，low-redundancy au tomated solution for early skin cancer diagnosis，demonstrating significant clinical practical value.

Key words： image classification； feature selection; InfoNCE loss；vision transformer model

1引言

皮肤癌在全球范围内的发病率呈逐年上升趋势，直接加剧了癌症相关死亡率的增长。（剩余18913字）