Application of transformer-enhanced convolutional neural network: multicenter MRI assessment of muscle invasion in bladder cancer

Abstract

Objective: Accurate preoperative assessment of muscle invasion in bladder cancer (BCa) guides therapy selection. However, MRI interpretation varies across readers and lesion morphologies. Therefore, we aimed to overcome the morphology-associated diagnostic bias through a deep learning method.

Materials and methods: This multicenter study included 1374 patients with BCa. An nnU-Net was fine-tuned to assist in lesion segmentation on T2-weighted images, providing inputs for a 2.5D ConvNeXt-tiny model to assess muscle invasion. The performance of the model was compared between pedunculated and sessile lesions. Furthermore, a head-to-head comparison was conducted among the model, a senior radiologist, and a junior radiologist.

Results: The validation Dice coefficient of nnU-net was 0.834. In the validation and three prospective test sets, the ConvNeXt-tiny model achieved areas under the receiver-operating characteristic curve of 0.915-0.925 for identifying muscle invasion in BCa, with accuracies of 84.9-91.0%, sensitivities of 81.3-96.2%, and specificities of 81.1-93.8%. In the subgroup analysis of pedunculated and sessile lesions, the model's diagnostic performance showed no significant difference across all datasets. In contrast, the two radiologists' specificities declined from around 90% in pedunculated lesions to approximately 75% (p = 0.010-0.050) in sessile lesions across the validation set, internal test set, and external test set 1. Therefore, in the head-to-head comparison of sessile lesions, the model demonstrated significantly higher specificities (91.9-96.0%) than the two radiologists (72.8-79.8%, p < 0.001-0.005) across the same datasets.

Conclusions: The DL model assesses muscle invasion in BCa independently of lesion morphology and holds potential for clinical application, particularly in sessile lesions.

Key points: Question Conventional MRI assessment of muscle invasion risk in pedunculated and sessile bladder cancers may be biased, but clear evidence and potential solutions are still lacking. Findings Morphology-associated diagnostic bias indeed exists, mainly as overstaging of sessile bladder cancer, while deep learning assessment of muscle invasion risk is morphology-independent. Clinical relevance The transformer-enhanced convolutional neural network developed in this study effectively reduces overestimation of muscle invasion risk in sessile bladder cancer and may serve as a complementary imaging tool for clinical evaluation.

Keywords: Bladder cancer; Convolutional neural network; MRI; Transformer; nnU-Net.