LGH-Transformer: A Hybrid CNN–Transformer Model for Multi-Class Bladder Tissue Classification from Cystoscopic Images

Abstract

Automated analysis of cystoscopic images may support consistent characterization of bladder tissue patterns, but reported performance is highly sensitive to patient composition, data partitioning, and evaluation methodology. Objective: This study provides an exploratory comparison of convolutional neural networks, Transformer-based models, and a proposed Local–Global Hybrid Transformer (LGH-Transformer) for four-class bladder tissue image classification. The dataset comprised 1754 frames and used the publicly released 1257/308/189 training–validation–held-out test split. The annotation-defined held-out test composition was HGC = 74, LGC = 53, NST = 37, and NTL = 25 (n = 189). Images were resized to 224 × 224 pixels, normalized, and augmented only in the training subset. Adam and RMSprop were evaluated. Accuracy and macro-averaged precision, recall, and F1-score were calculated from the integer-count evaluation matrices. The highest held-out accuracy, 187/189 (0.9894), was reached by seven model–optimizer configurations: DenseNet201–RMSprop, XCiT Medium 24 P16–Adam, ConvNeXt Base–Adam, CoAtNet–Adam, ConvNeXt Base–RMSprop, CoAtNet–RMSprop, and the proposed LGH-Transformer–RMSprop; MobileNet–Adam and Xcep-tion–RMSprop each yielded 185/189 (0.9788), and LGH-Transformer–Adam yielded 186/189 (0.9841). On a 189-frame test set, differences in one to two frames do not establish statistical superiority, so these configurations are reported as a top-accuracy tier rather than as a ranked superiority claim. Within this tier, the LGH-Transformer had the smallest computational footprint (≈18.2 M parameters and ≈4.1 GFLOPs, versus 20.0–88.6 M parameters and 4.3–16.2 GFLOPs for the other top-tier models). Component-wise ablation experiments showed accuracy decreases of 0.0106, 0.0053, and 0.0053 after removing the CNN tokenizer, local attention, and global attention, respectively. The LGH-Transformer reaches the top accuracy tier of this 18-architecture benchmark at the lowest parameter and FLOP budget within that tier, supporting an internally consistent, efficiency-oriented comparison within the available cohort. The absence of task-aligned external validation and the small number of held-out patients require the findings to be interpreted as exploratory internal validation rather than evidence of clinical generalizability.