Diffusion-Based Histopathology Enhancement with Morphology-Aware Deep Learning for Accurate Lung Cancer Detection

Abstract The early detection of lung cancer through histopathological images is still a major challenge, mainly because of intrinsic noise, different staining effects, and the very subtle microscopic changes that differentiate benign from malignant tissues. These issues are a cause of reduced classification reliability, especially in automated diagnostic systems. This work is focused on designing a morphology-aware deep learning system augmented by a diffusion-based structural enhancement method to deliver more accurate classification results that are also more explainable in the context of lung histopathological image analysis. Our proposed framework's performance was checked using the LC25000 histopathology dataset—a collection of 25,000 images depicting benign lung tissue as well as three major types of lung cancer, including adenocarcinoma, squamous cell carcinoma, and large cell carcinoma. Anisotropic diffusion was used to remove structural noise, but at the same time, maintain the morphological details that are important to the diagnosis. The deep learning network, which integrates hierarchical feature extraction and attention-based fusion, was then used to classify these processed images. The proposed method's performance was benchmarked against well-known deep architectures such as ResNet50, DenseNet121, Vision Transformer, and EfficientNet, and B4. The results of image enhancement were quantified through PSNR (peak signal-to-noise ratio) and SSIM figures. The experimental findings showed that the proposed framework outclassed other competing models in terms of accuracy (97.84%), F1-score (97.34%), and area under the curve (0.982). The diffusion enhancement method not only elevated the perceptual quality of the images but also drastically improved the metrics where the PSNR rose from 24.18 to 31.72 dB and the SSIM went up from 0.78 to 0.92. The presented framework combining morphology-aware learning and diffusion-based structural augmentation manages to improve not only the classification performance but also the interpretability. The study evidences that the method has great potential for dependable computer-aided lung disease diagnosis.