融合多维依赖建模与空间自适应特征学习的 皮肤病变分类方法

Skin Lesion Classification via Integrated Multi-Dimensional Dependency Modeling and Spatial Adaptive Feature Learning

摘要: 皮肤癌作为全球致死率极高的恶性肿瘤之一，其早期精准分类对于临床辅助诊断与治疗至关重要。针对现有卷积神经网络在处理形态多变、大小不一的皮肤病损图像时，存在多尺度特征捕获不足以及对辨别细粒度病灶聚焦不充分等挑战，本文提出了一种融合分层上下文聚合与多核特征协同学习的改进型深度学习分类框架。首先，针对病灶区域多尺度的特性，设计了分层上下文聚合模块(HCAM)，通过空间金字塔聚合与深度细化机制，动态提取并融合跨尺度的全局上下文特征。其次，构建了多核特征融合模块(MKFFM)，利用多路并行且感受野递增(1 × 1至7 × 7)的深度可分离卷积，自适应地捕捉不同尺度的局部纹理特征，突破了单一卷积核信息提取的局限性。此外，本文引入了基于空间注意力引导的全局–局部双路特征聚合机制，并在训练阶段结合CutMix混合数据增强策略，有效抑制了复杂皮肤背景的干扰，显著增强了模型在数据不平衡及小样本场景下的泛化能力与鲁棒性。基于ISIC 2018等公开数据集的五折交叉验证实验结果表明，该模型显著提升了各类皮肤病变的分类精确度与F1值，为计算机辅助皮肤病诊断提供了一种兼具高鲁棒性与解释性的技术方案。

Abstract: As one of the most fatal malignant tumors globally, the early and accurate classification of skin cancer is crucial for clinical auxiliary diagnosis and treatment. To address the challenges of insufficient multi-scale feature capture and inadequate focus on discriminative fine-grained lesions in existing convolutional neural networks when processing skin lesion images with variable morphologies and sizes, this paper proposes an improved deep learning classification framework integrating hierarchical context aggregation and multi-kernel feature collaborative learning. First, targeting the multi-scale characteristics of lesion regions, a Hierarchical Context Aggregation Module (HCAM) is designed to dynamically extract and fuse cross-scale global context features through spatial pyramid aggregation and depthwise refinement mechanisms. Second, a Multi-Kernel Feature Fusion Module (MKFFM) is constructed, which utilizes multi-branch parallel depthwise separable convolutions with increasing receptive fields (from 1 × 1 to 7 × 7) to adaptively capture local texture features at different scales, breaking through the limitations of single convolution kernel information extraction. Furthermore, this paper introduces a global-local dual-branch feature aggregation mechanism guided by spatial attention, and combined with the CutMix mixed data augmentation strategy during the training phase, it effectively suppresses the interference of complex skin backgrounds and significantly enhances the generalization ability and robustness of the model under data imbalance and small-sample scenarios. The five-fold cross-validation experimental results based on public datasets such as ISIC 2018 demonstrate that the proposed model significantly improves the classification accuracy and F1-score of various skin lesions, providing a highly robust and interpretable technical solution for computer-aided dermatology diagnosis.