Quality Controlled Synthetic Data Generation with LoRA-Adapted Stable Diffusion for Transformer-Based Lung Cancer...

The classification of lung cancer subtypes from histopathological images remains a challenging problem for deep learning models due to the limited amount of data and the intraclass morphological diversity. In this study, the impact of synthetic histopathology images generated solely from training images on classification performance was investigated within the integrated multistage QCL-DiffSynTrans framework on the LungHist700 dataset. In synthetic data generation, the Stable Diffusion model, class-specific LoRA adaptations, and the image-to-image generation approach were used. To enhance color harmony, LAB color matching was applied to the generated images based on the source training images. The final Img2Img Synthetic 25 dataset consists of 121 synthetic images; these images were only added to the training set, while the validation and test sets were completely preserved with real images. In the selection of synthetic images, filtering steps such as duplicate, source-copy, severe color/QC inconsistency, DINOv2 similarity check, and dual-model label consistency were applied. The final synthetic dataset has passed all the specified quality controls, and the obtained results have been reported in the study. In the classification phase, five independent ViT-B/16 models were trained, and the five-seed probability ensemble approach was used by averaging the output probabilities of these models for the final decision. Compared to the real-only ensemble, the Real + Img2Img Synthetic 25 ensemble approach increased the accuracy value from 0.9327 to 0.9519 and the macro-F1 value from 0.9363 to 0.9558. MCC increased from 0.8962 to 0.9263, supporting more consistent agreement between the ensemble predictions and the reference labels. Macro ROC-AUC was consistently high at 0.9876 across both training conditions, indicating class separability well above chance and preserved discriminative performance; the ensemble-level gains were therefore interpreted descriptively. These findings indicate that img2img-based synthetic histopathology images, derived solely from training images and subjected to quality control, can provide a measurable contribution to the classification within the QCL-DiffSynTrans framework. Results show that controlled selection of synthetic images may be a useful data augmentation strategy for limited histopathology datasets. The proposed approach is significant in terms of building more robust and reliable digital pathology decision support systems.