Automatic lung cancer classification using histopathological images: A hybrid DenseNet121-slantlet transform framework

Abstract

Accurate classification of histopathological images is essential in the early diagnosis and treatment of lung cancer. Conventional deep learning approaches often face the challenges of simulating fine-grained texture features of low-level features as well as high-level semantic features that can be observed in complex tissue structures. To address this issue, the current paper introduces a novel hybrid deep learning architecture that combines SLT and DenseNet121 for accurate lung cancer classification. Publicly available datasets and clinical settings are used to preprocess and diversify histopathological images in real time, reducing noise and enhancing image quality. DenseNet121 extracts rich hierarchical spatial features, whereas SLT captures discriminative multi-resolution frequency-domain texture features. These complementary features are combined on the feature level and grouped by a Softmax classifier. The grad-weighted Class Activation Mapping (Grad-CAM) is employed to highlight the regions contributing to the diagnosis, thereby improving model interpretability. The effectiveness of the proposed approach is demonstrated by a set of experiments, such as k-fold cross-validation, ablation studies, and out-of-sample validation of The Cancer Genome Atlas data. In this regard, the proposed DenseNet121–SLT framework records a classification accuracy of 98.78% and an AUC of 0.994, which is robust and generalizable. Further clinical validation on larger multi-centre datasets is required before clinical deployment.