Shanmugam Karthikeyan, Rajaguru Harikumar
Bannari Amman Institute of Technology, Tamil Nadu 638401, India.
Diagnostics (Basel). 2025 Mar 21;15(7):805. doi: 10.3390/diagnostics15070805.
Lung cancer is a leading cause of cancer-related mortalities, with early diagnosis crucial for survival. While biopsy is the gold standard, manual histopathological analysis is time-consuming. This research enhances lung cancer diagnosis through deep learning-based feature extraction, fusion, optimization, and classification for improved accuracy and efficiency. The study begins with image preprocessing using an adaptive fuzzy filter, followed by segmentation with a modified simple linear iterative clustering (SLIC) algorithm. The segmented images are input into deep learning architectures, specifically ResNet-50 (RN-50), ResNet-101 (RN-101), and ResNet-152 (RN-152), for feature extraction. The extracted features are fused using a deep-weighted averaging-based feature fusion (DWAFF) technique, producing ResNet-X (RN-X)-fused features. To further refine these features, particle swarm optimization (PSO) and red deer optimization (RDO) techniques are employed within the selective feature pooling layer. The optimized features are classified using various machine learning classifiers, including support vector machine (SVM), decision tree (DT), random forest (RF), K-nearest neighbor (KNN), SoftMax discriminant classifier (SDC), Bayesian linear discriminant analysis classifier (BLDC), and multilayer perceptron (MLP). A performance evaluation is performed using K-fold cross-validation with K values of 2, 4, 5, 8, and 10. The proposed DWAFF technique, combined with feature selection using RDO and classification with MLP, achieved the highest classification accuracy of 98.68% when using K = 10 for cross-validation. The RN-X features demonstrated superior performance compared to individual ResNet variants, and the integration of segmentation and optimization significantly enhanced classification accuracy. The proposed methodology automates lung cancer classification using deep learning, feature fusion, optimization, and advanced classification techniques. Segmentation and feature selection enhance performance, improving diagnostic accuracy. Future work may explore further optimizations and hybrid models.
肺癌是癌症相关死亡的主要原因之一，早期诊断对生存至关重要。虽然活检是金标准，但手动组织病理学分析耗时较长。本研究通过基于深度学习的特征提取、融合、优化和分类来提高肺癌诊断的准确性和效率。该研究首先使用自适应模糊滤波器进行图像预处理，然后使用改进的简单线性迭代聚类（SLIC）算法进行分割。分割后的图像被输入到深度学习架构中，具体为ResNet-50（RN-50）、ResNet-101（RN-101）和ResNet-152（RN-152），以进行特征提取。使用基于深度加权平均的特征融合（DWAFF）技术对提取的特征进行融合，生成ResNet-X（RN-X）融合特征。为了进一步优化这些特征，在选择性特征池化层中采用了粒子群优化（PSO）和马鹿优化（RDO）技术。使用包括支持向量机（SVM）、决策树（DT）、随机森林（RF）、K近邻（KNN）、SoftMax判别分类器（SDC）、贝叶斯线性判别分析分类器（BLDC）和多层感知器（MLP）在内的各种机器学习分类器对优化后的特征进行分类。使用K值为2、4、5、8和10的K折交叉验证进行性能评估。所提出的DWAFF技术，结合使用RDO进行特征选择和使用MLP进行分类，在使用K = 10进行交叉验证时达到了98.68%的最高分类准确率。RN-X特征与单个ResNet变体相比表现出卓越的性能，并且分割和优化的整合显著提高了分类准确率。所提出的方法使用深度学习、特征融合、优化和先进的分类技术实现了肺癌分类的自动化。分割和特征选择提高了性能，提升了诊断准确性。未来的工作可能会探索进一步的优化和混合模型。
