膀胱镜是膀胱肿瘤的重要检查方法，但是检查效果受病变特点、检查者水平等影响大，容易出现误诊、漏诊。深度学习模型的应用将有助于提高膀胱镜下肿瘤的识别能力。本研究将探讨深度学习模型对膀胱镜下不同大小、不同清晰度肿瘤的分割性能。

回顾性纳入北京协和医院、北京市第六医院和北京隆福医院2022年6月1日至2024年6月1日接受经尿道膀胱肿瘤切除术的患者共198例。应用语义分割模型Mask2Former和DeepLabV3+对膀胱镜图像中的肿瘤进行分割。分别以171例和27例患者的视频资料作为训练集和测试集。观察指标包括交并比、准确率、召回率和平均Dice系数。应用R 4.0.2软件进行分析，所有的计数资料均以绝对数或百分比进行表示。

DeepLabV3+模型（参数量21.75 M）对低清晰度膀胱肿瘤的分割准确率和召回率均为83.3%；Mask2Former模型（参数量197 M）对低清晰度膀胱肿瘤的分割准确率和召回率分别为92.1%和88.9%；两种模型对高清晰度膀胱肿瘤的分割准确率和召回率均高于98.0%。当应用Mask2Former模型时，随着训练集病例由10例增加至171例，模型对膀胱肿瘤的分割准确率由94.0%升高至97.7%、召回率由73.3%升高至93.5%，mDice由62.3%升高至85.5%，其中低清晰度组的提升尤为明显。应用Mask2Former模型对不同大小膀胱肿瘤进行分割，肿瘤占视野面积<0.02组与>0.1组的准确率分别为90.9%和98.1%；召回率分别为83.3%和99.8%。

深度学习模型Mask2Former对膀胱镜下膀胱肿瘤的分割性能良好，尤其对低清晰度、肿瘤占视野面积大的图像具有较好的分割能力。

Cystoscopy is an essential diagnostic procedure for bladder tumors. However, the effectiveness of the examination is greatly influenced by the characteristics of the lesion and the proficiency of the examiner, leading to possible misdiagnosis and missed diagnosis. The application of deep learning models can help improve the identification of tumors under cystoscopy. This study explores the performance of deep learning models in segmenting tumors of different sizes and clarity under cystoscopy.

A retrospective study included 198 patients who underwent transurethral bladder tumor resection in Peking Union Medical College Hospital, Beijing Sixth Hospital, and Beijing Longfu Hospital from June 1, 2022, to June 1, 2024. Semantic segmentation models Mask2Former and DeepLabV3+ were used to identify bladder tumors. Video data from 171 patients and 27 patients were used as the training set and test set, respectively. Observational indicators included intersection over union (IOU), accuracy, recall rate, and mean Dice coefficient. Analysis was performed using R 4.0.2 software, and all count data were expressed as numbers or percentages.

The DeepLabV3+ model (with 21.75 Mparameters) achieved an accuracy and recall of 83.3% for segmenting low-clarity bladder tumors. The Mask2Former model (with 197 M parameters) achieved an accuracy and recall rate of 92.1% and 88.9%, respectively, for low-clarity bladder tumors. Both models achieved an accuracy and recall rate of over 98.0% for high-clarity bladder tumors. When applying the Mask2Former model, as the number of cases in the training set increased from 10 to 171, the model's accuracy rate for segmenting bladder tumors increased from 94.0% to 97.7%, recall rate increased from 73.3% to 93.5%, and mDice increased from 62.3% to 85.5%, with a particularly noticeable improvement in the low-clarity group. When using the Mask2Former model to segment bladder tumors of different sizes, the accuracy rate for tumors occupying less than 0.02 of the screen and more than 0.1 was 90.9% and 98.1%, respectively; the recall rate was 83.3% and 99.8%, respectively.

The deep learning model Mask2Former demonstrates excellent performance in the segmentation of bladder tumors under cystoscopy, particularly exhibiting robust segmentation capabilities for low-clarity images and tumors occupying a significant portion of the visual field.