背景：

在癌症研究中，早期精准诊断和个体化治疗是改善患者预后的关键。尽管免疫治疗显著改变了癌症治疗模式，但免疫无应答现象仍然普遍存在。糖酵解作为肿瘤细胞的重要代谢途径，与免疫逃逸密切相关，并可能影响免疫治疗的疗效。膀胱癌是常见的泌尿系统肿瘤，在分子亚型及临床疗效方面具有高度异质性。此外，膀胱肿瘤的多发性和高复发率增加了膀胱镜检查的漏诊风险，并可能导致检查结果的不一致性，亟需更精准的系统化策略优化诊疗模式。随着人工智能及单细胞转录组测序技术的快速发展，医学研究已具备对肿瘤特征进行深度和高分辨率分析的能力。本研究基于大规模多组学数据分析，并结合多种人工智能算法，系统探讨糖酵解与免疫治疗反应的关系，同时鉴定膀胱癌的分子分型，以推动个体化治疗方案的实施。同时，为提升膀胱镜检查的准确性，我们进一步引入深度学习模型HRNetV2，评估其在膀胱镜下肿瘤智能识别中的应用潜力，以期构建综合性的膀胱癌诊断和治疗预测体系。

方法：

为了探究糖酵解与免疫治疗反应的相关性，本研究基于34个单细胞转录组测序数据集进行泛癌分析，构建糖酵解基因特征（Glycolysis.Sig），并结合大规模泛癌队列和免疫检查点抑制剂治疗转录组数据集，评估其对免疫疗效的预测能力。进一步采用机器学习算法筛选关键糖酵解基因特征（Hub-Glycolysis.Sig），并评估其在临床预后模型中的表现。通过一致性聚类分析鉴定膀胱癌分子亚型，评估亚型间免疫浸润及突变特征差异，结合膀胱癌样本单细胞转录组测序验证组织中的糖酵解水平差异。在膀胱病变智能识别方面，本研究收集94例患者的102段白光膀胱镜视频，并按4：1比例划分为训练集和测试集，采用HRNetV2语义分割模型识别膀胱病变形态特征。测试集根据图像辨识度划分为高辨识度组和低辨识度组，以分析影响模型识别效果的图像质量因素。通过灵敏度、精确率和平均Dice分数（mDice）评估模型性能。

结果：

对两个免疫治疗单细胞测序队列的分析表明糖酵解与免疫治疗反应密切相关。糖酵解基因特征（Glycolysis.Sig）在泛癌数据集中表现出卓越的免疫疗效预测能力。关键糖酵解基因特征（Hub-Glycolysis.Sig）在不同癌症类型中与免疫微环境的相互作用存在差异。在膀胱癌中，高糖酵解风险评分与不良预后相关。不同膀胱亚型具有不同的免疫浸润模式和突变特征。单细胞测序结果显示，膀胱癌组织中上皮细胞具有较高的糖酵解水平。此外，通过CRISPR筛选、药物敏感性分析和分子对接，鉴定出多个潜在基因靶点。膀胱病变检测部分，共标注33657帧图像，测试集中HRNetV2模型总体灵敏度和精确率分别为91.6%和91.3%，mDice分数为80.3%。在高辨识度组中，灵敏度和精确率分别为94.8%和94.4%，低辨识度组则分别为75.6%和74.8%。分析模型漏检原因主要包括目标距离近、体积小及形态非典型；误检则多由于特征相似性高或小目标识别偏差所致。

结论：

本研究基于大规模数据分析验证了糖酵解与免疫治疗反应的负相关性，并构建了糖酵解基因特征（Glycolysis.Sig）作为新型预测生物标志物，具有良好的临床应用潜力。此外，关键糖酵解基因特征（Hub-Glycolysis.Sig）可用于膀胱癌亚型鉴定，为个体化治疗提供新思路。后续将进一步验证糖酵解相关基因的生物学功能并深入探索其分子机制。在膀胱病变智能检测方面，HRNetV2模型表现优异，展现出较高的灵敏度和精确率。未来研究将优化算法，并在更大规模的多中心数据集中验证其临床应用潜力。

Background

Early cancer diagnosis and personalized precision therapy are critical challenges in oncology research. Although immunotherapy has revolutionized cancer treatment, immune resistance remains a prevalent issue. Glycolysis, a key metabolic pathway in tumor cells, is closely associated with immune evasion. Bladder cancer, a common malignancy of the urinary system, exhibits high molecular heterogeneity, clonal diversity, and variable clinical outcomes. Additionally, its multifocal nature and high recurrence rate increase the risk of missed diagnoses during cystoscopy, potentially leading to inconsistent diagnostic outcomes. Therefore, a more precise and systematic approach is urgently needed to optimize diagnostic and therapeutic strategies. In this study, we conducted large-scale data analysis combined with artificial intelligence algorithms to systematically investigate the relationship between glycolysis and immunotherapy response while identifying molecular subtypes of bladder cancer to facilitate personalized treatment strategies. Furthermore, we applied deep learning techniques to evaluate the performance of the HRNetV2 model in detecting bladder lesions, aiming to improve early detection.

Methods

To explore the association between glycolysis and immunotherapy response, we performed a pan-cancer analysis across 34 single-cell RNA sequencing (scRNA-seq) datasets and constructed Glycolysis.Sig. Using large-scale pan-cancer cohorts and immune checkpoint inhibitor (ICI) treatment RNA-seq datasets, we assessed its predictive potential for immunotherapy response. Machine learning algorithms were further employed to identify Hub-Glycolysis.Sig and evaluate its performance in clinical prognostic models. Consensus clustering analysis was conducted to classify molecular subtypes of bladder cancer, followed by an assessment of immune infiltration and mutational characteristics across subtypes. scRNA-seq analysis of bladder cancer samples was performed to validate glycolysis level differences within tumor tissues. For intelligent bladder lesion detection, we collected 102 white light cystoscopy videos from 94 patients and divided them into training and test sets in a 4:1 ratio. The HRNetV2 semantic segmentation model was employed to detect bladder lesions, and the test set was categorized into high- and low-resolution groups based on image quality. Model performance was evaluated using sensitivity, precision, and the mean Dice coefficient (mDice).

Results

Analysis of two ICI-treated scRNA-seq cohorts revealed a potential association between glycolysis and immunotherapy response. Glycolysis.Sig demonstrated robust predictive capability for immunotherapy response across pan-cancer datasets. The key glycolysis-related gene signature, Hub-Glycolysis.Sig, exhibited distinct interactions with the immune microenvironment across different cancer types. In bladder cancer, a high glycolysis risk score correlated with poor prognosis. Molecular subtyping identified distinct immune infiltration patterns and mutational landscapes among bladder cancer subtypes. scRNA-seq analysis confirmed elevated glycolysis levels in bladder epithelial cells. In bladder lesion detection, a total of 33,657 frames were annotated, with the test set achieving an overall sensitivity of 91.6%, precision of 91.3%, and an mDice score of 80.3%. The high-resolution group demonstrated a sensitivity of 94.8% and precision of 94.4%, whereas the low-resolution group achieved 75.6% and 74.8%, respectively. False positives were primarily due to abnormal mucosal texture features and the misidentification of small targets as lesions. False negatives were attributed mainly to atypical target features, relatively small target sizes, and targets that were too close to the observation point.

Conclusion

This study validated the negative correlation between glycolysis and immunotherapy response through large-scale data analysis and established Glycolysis.Sig as a novel predictive biomarker with promising clinical applicability. Additionally, the key glycolysis-related gene signature, Hub-Glycolysis.Sig, was identified as a potential tool for molecular subtyping of bladder cancer, providing new insights for personalized treatment. Future research will further explore the biological functions and molecular mechanisms of glycolysis-related genes. In bladder lesion detection, HRNetV2 demonstrated superior performance with high sensitivity and precision. Further optimization and validation in larger multi-center cohorts will be conducted to assess its clinical utility.