Guo Xiaoyan, Chen Jiyue, Lin Pingju, Lu Qi, Kou Ting, Li Kun, Yang Shiming, Shen Weidong
Senior Department of Otolaryngology Head and Neck Surgery, the 6th Medical Center of Chinese PLA General Hospital & Chinese PLA Medical School National Clinical Research Center for Otolaryngologic Diseases, State Key Laboratory of Hearing and Balance Science, Beijing, China.
Institute of Interdisciplinary Medicine and Engineering, University of Southern California, Keck School of Medicine, Los Angeles, California, America.
Acta Otolaryngol. 2025 Aug;145(8):759-769. doi: 10.1080/00016489.2025.2502562. Epub 2025 May 8.
While accurate evaluation of PFP is crucial for determining optimal treatment strategies, current clinical assessments rely heavily on subjective evaluations, leading to considerable variability between inter- and intra-observer ratings.
This study aimed to develop and validate a collaborative framework for evaluating PFP based on regions of interest in facial images.
We developed and tested two approaches: (1) a collaborative framework integrating image interpretation techniques (representation learning CNN) with predefined handcrafted features based on regions of interest in facial images, and (2) a convolutional neural network (CNN) model trained exclusively on full-face patient images. The diagnostic accuracy of both systems was evaluated using a test set and compared with otologists' assessments.
The collaborative framework achieved a mean Area Under the Curve (AUC) of 0.92 for PFP prediction in the test set, surpassing the 0.76 AUC achieved by the CNN trained on full-face images. The framework's performance matched that of experienced otologists (accuracy: 80.0% vs. 77.2%; sensitivity: 85.3% vs. 77.7%). Moreover, system assistance improved primary clinicians' mean accuracy by 17.7 percentage points.
These findings demonstrate that our collaborative framework-based automated diagnosis system can effectively assist clinicians in PFP diagnosis.
虽然准确评估镫骨足板固定（PFP）对于确定最佳治疗策略至关重要，但目前的临床评估在很大程度上依赖主观评估，导致观察者间和观察者内评分存在相当大的差异。
本研究旨在开发并验证一种基于面部图像感兴趣区域评估PFP的协作框架。
我们开发并测试了两种方法：（1）一种协作框架，将图像解释技术（表征学习卷积神经网络）与基于面部图像感兴趣区域的预定义手工特征相结合；（2）一种仅在全脸患者图像上训练的卷积神经网络（CNN）模型。使用测试集评估这两种系统的诊断准确性，并与耳科医生的评估结果进行比较。
协作框架在测试集中对PFP预测的平均曲线下面积（AUC）为0.92，超过了在全脸图像上训练的CNN所达到的0.76的AUC。该框架的性能与经验丰富的耳科医生相当（准确率：80.0%对77.2%；灵敏度：85.3%对77.7%）。此外，系统辅助使初级临床医生的平均准确率提高了17.7个百分点。
这些发现表明，我们基于协作框架的自动诊断系统可以有效地协助临床医生进行PFP诊断。
