Abstract: Objective The aim of this study was to investigate whether a multimodal artificial intelligence diagnostic model can be constructed based on ultrasound images and urinary cytology to improve diagnostic sensitivity and assist in reducing the use of cystoscopy. Methods This was a single-center retrospective study that included 2056 patients who underwent both bladder ultrasound examination and urinary cytology examination from January 2018 to September 2023 for model training and validation. The gold standard was determined based on the patients' histopathological results, and patients with negative results needed to be followed up for 6 months to confirm non-cancer. Firstly, we constructed an AI diagnostic model for bladder cancer based on ResNet model and ultrasound images. We used pre-trained weights on ImageNet as the initialization of model weights. Random gradient descent and cross-entropy loss were used for network weight adjustment and algorithm optimization. After the ultrasound AI model output the diagnostic score, we combined it with the results of urinary cytology diagnosis and clinical risk factors based on Logistic regression to construct a multimodal diagnostic model, and output the final diagnostic probability for each individual. The effectiveness of the model was then validated in the validation set and subgroups （including different stages, grades, and clinical scenarios）. The final multimodal model was named BCaUSNet. Results The BCaUSNet model had a diagnostic sensitivity of 0.896（95%CI： 0.839-0.938） and an area under the curve of 0.917（95%CI： 0.891-0.942） in the validation set. In the scenario of recurrence monitoring, the model's sensitivity could reach 0.821（95%CI： 0.631-0.939）, and the negative predictive value could reach 0.896（95%CI： 0.773-0.965）, which could assist in reducing the use of cystoscopy with a high degree of certainty. In tumors with low malignant potential and low-grade tumors where urinary cytology is difficult to diagnose, the BCaUSNet model increased their sensitivity to 71.4% and 93.3%, respectively. In non-muscular invasive tumors and small tumors （<1.5 cm） that are easily missed by ultrasound, the BCaUSNet model increased their sensitivity to 89.5% and 87.5%, respectively. Conclusion The construction of a multimodal artificial intelligence diagnostic model for bladder cancer based on ultrasound images and urinary cytology has high diagnostic sensitivity, which helps to reduce missed diagnosis of bladder cancer, reduce the use of cystoscopy, and has good clinical utility and innovative significance.

Key words: bladder cancer, ultrasound imaging, urinary cytology, artificial intelligence, multi-modal diagnostic model

摘要： 目的 本研究旨在探究是否可以基于超声图像及尿液细胞学构建多模态人工智能诊断模型,提升诊断敏感性,并辅助减少膀胱镜的使用。方法 本研究是一项单中心回顾性研究,纳入了从2018年1月至2023年9月份同时行膀胱超声检查及尿液细胞学检查的2056名患者进行模型训练及验证。金标准基于患者的组织病理结果进行确定,阴性患者需要进行6个月的随访以确定未患癌。基于ResNet（残差神经网络）模型以及超声影像构建了膀胱癌人工智能（AI）诊断模型,使用在ImageNet数据集上预训练的权重作为模型权重的初始化,并采用随机梯度下降和交叉熵损失进行网络权重调整和算法优化,当超声AI模型输出诊断评分后,将其与尿液细胞学诊断结果及临床危险因素基于Logistic回归构建多模态诊断模型并为每个个体输出最终诊断概率。最终在验证集以及亚组（包括不同分期、分级、临床场景）中验证模型的有效性。最终的多模态模型将命名为BCaUSNet（膀胱癌超声残差神经网络模型）。结果 BCaUSNet模型在验证集中的诊断敏感性为0.896（95%CI：0.839~0.938）,曲线下面积为0.917（95%CI：0.891~0.942）。在复发监测场景中,模型的敏感性可达到0.821（95%CI：0.631~0.939）,阴性预测值可达到0.896（95%CI：0.773~0.965）,这可以在有较高把握的程度上辅助减少60%的膀胱镜使用。在尿液细胞学难以诊断的低度恶性潜能肿瘤及低级别肿瘤中,BCaUSNet模型敏感性分别提高至71.4%及93.3%。在超声较易漏诊的非肌层浸润性肿瘤以及小肿瘤（<1.5 cm）中,BCaUSNet模型敏感性分别提高至89.5%及87.5%。结论 基于超声图像及尿液细胞学构建膀胱癌多模态人工智能诊断模型具有较高的诊断敏感性,有助于降低膀胱癌的漏诊、减少膀胱镜的使用,具有较好的临床效用以及创新意义。

关键词: 膀胱癌, 超声影像, 尿液细胞学, 人工智能, 多模态诊断模型