文献速递泌尿生殖技术与安全CT

优化深度学习框架用于CT肾结石检测

A fully optimized deep learning framework for explainable and efficient kidney stone detection in computed tomography imaging.
PubMed 文献 · Abdom Radiol (NY) · 2026-06-27

研究开发了一种经超参数优化的自定义卷积神经网络,用于轴位CT肾结石检测,并与六种预训练骨干网络在统一贝叶斯优化条件下比较。主要结论为自定义模型在测试准确率0.9960、F1 0.9957上与ResNet101持平,但推理延迟最低且参数量最少。外部验证显示20%域内校准后F1恢复至0.9806。

要点

  • 自定义模型和ResNet101测试表现并列最佳,准确率0.9960,F1 0.9957,ROC-AUC≥0.9995。
  • 自定义模型单样本推理延迟2.24 ms,比基线快1.5–53倍,参数量1.53 M,少2.3–28倍。
  • Grad-CAM++激活区域解剖位置合理,有助于提示模型关注结石相关区域。

核心结论

  • 统一的贝叶斯协议下自定义架构搜索能以更低的计算成本匹配充分优化的预训练骨干网络的性能上限。
  • 方法学上一致的基准比较而非仅报告峰值准确率对于临床可部署的肾结石检测至关重要。

研究设计与样本

  • 开发自定义残差卷积神经网络,训练集3,364张轴位CT切片,其中结石1,577张,非结石1,787张,按70/15/15分层划分。
  • 对自定义模型和ResNet101、DenseNet121、EfficientNetB3、MobileNetV3Large、VGG16、Xception六种预训练网络统一采用贝叶斯超参数优化,每种架构1,000次试验。
  • 采用分层五折交叉验证、十次固定划分重复运行以及95%置信区间评估稳健性,并使用独立的孟加拉国CT队列4,121张切片进行外部验证。

关键影像表现/征象

  • Grad-CAM++生成的激活区域在解剖上合理,指向与肾结石相关的CT区域。

诊断标准与关键数值

  • 自定义模型与ResNet101在测试集上准确率均为0.9960,F1均为0.9957,ROC-AUC均≥0.9995。
  • 自定义模型单样本推理延迟为2.24 ms,比基线模型快1.5–53倍;参数量为1.53 M,比基线少2.3–28倍。
  • 五折交叉验证Macro-F1为0.9856±0.0036,Friedman检验显示不同架构间存在显著差异,P<0.001。
  • 外部孟加拉国CT队列上,20%域内校准后F1恢复至0.9806。

临床应用与阅片要点

  • 在肾结石CT辅助检测中,自定义轻量模型可在保证准确率的同时显著降低推理延迟和参数量,利于临床部署。
  • 评估深度学习模型时应关注统一优化条件下的多指标比较和外部验证表现,而非仅看单一准确率。

局限性

  • 外部验证队列的初始表现受域偏移影响,需经20%域内校准后F1才恢复至0.9806。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

英文原文(摘要)· 点击展开

PURPOSE: To develop and rigorously evaluate a hyperparameter-optimized custom convolutional neural network for kidney stone detection on axial computed tomography (CT) and to compare it, under matched optimization conditions, with widely used pretrained backbones, emphasizing statistical reliability, computational efficiency, and cross-institutional generalization.

METHODS: A custom convolutional neural network with residual blocks was trained on 3,364 axial CT slices (Stone n = 1,577; Non-stone n = 1,787) under a stratified 70/15/15 split. The same Bayesian hyperparameter optimization (1,000 trials per architecture, tree-structured Parzen estimator) was applied uniformly to the proposed model and six pretrained backbones (ResNet101, DenseNet121, EfficientNetB3, MobileNetV3Large, VGG16, Xception). Robustness was characterized by stratified five-fold cross-validation, ten repeated fixed-split runs with controlled seeds, and 95% confidence intervals. External validation used an independent Bangladesh CT cohort (4,121 slices). Interpretability was probed with Grad-CAM++.

RESULTS: The proposed model and ResNet101 jointly achieved the best test performance (accuracy 0.9960, F1 0.9957, ROC-AUC ≥ 0.9995). The proposed model attained the lowest single-sample inference latency (2.24 ms; 1.5-53× faster than baselines) with 1.53 M parameters (2.3-28× fewer). Five-fold cross-validation gave Macro-F1 0.9856 ± 0.0036; the Friedman test confirmed significant cross-architecture differences (p < 0.001). 20% in-domain calibration recovered F1 to 0.9806 on the external cohort. Grad-CAM + + activations were anatomically plausible.

CONCLUSION: Custom architecture search under a unified Bayesian protocol can match the ceiling of fully-optimized pretrained backbones at a fraction of their computational cost. Methodologically consistent benchmarking, not peak-accuracy reporting, is essential for clinically deployable kidney stone detection.

来源:Abdom Radiol (NY),查看原文。版权归原作者所有。

引用

  1. Caskurlu, Bakır, Zontul, et al. A fully optimized deep learning framework for explainable and efficient kidney stone detection in computed tomography imaging. Abdom Radiol (NY). 2026. PMID: 42362935. doi:10.1007/s00261-026-05640-9

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。