Toprak Serdar Ferit, Dedeoğlu Serkan, Kozan Günay, Ayral Muhammed, Can Şermin, Türk Ömer, Akdağ Mehmet
Department of Audiology, Artuklu University, Mardin 47100, Turkey.
Department of Otorhinolaryngology, University of Health Sciences Gazi Yasargil Training and Research Hospital, Diyarbakır 21100, Turkey.
Bioengineering (Basel). 2025 Aug 8;12(8):854. doi: 10.3390/bioengineering12080854.
Mucormycosis is a life-threatening fungal infection, where rapid diagnosis is critical. We developed a deep learning approach using paranasal computed tomography (CT) images to test whether mucormycosis can be detected automatically, potentially aiding or expediting the diagnostic process that traditionally relies on biopsy.
In this retrospective study, 794 CT images (from patients with mucormycosis, nasal polyps, or normal findings) were analyzed. Images were resized and augmented for training. Two transfer learning models (ResNet50 and ConvNeXt Small) were fine-tuned to classify images into the three categories. We employed a 70/30 train-test split (with five-fold cross-validation) and evaluated performance using accuracy, precision, recall, F1-score, and confusion matrices.
The ConvNeXt Small model achieved 100% accuracy on the test set (precision/recall/F1-score = 1.00 for all classes), while ResNet50 achieved 99.16% accuracy (precision ≈0.99, recall ≈0.99). Cross-validation yielded consistent results (ConvNeXt accuracy 99% across folds), indicating no overfitting. An ablation study confirmed the benefit of transfer learning, as training ConvNeXt from scratch led to lower accuracy (85%) Conclusions: Our findings demonstrate that deep learning models can accurately and non-invasively detect mucormycosis from CT scans, potentially flagging suspected cases for prompt treatment. These models could serve as rapid screening tools to complement standard diagnostic methods (histopathology), although we emphasize that they are adjuncts and not replacements for biopsy. Future work should validate these models on external datasets and investigate their integration into clinical workflows for earlier intervention in mucormycosis.
毛霉菌病是一种危及生命的真菌感染，快速诊断至关重要。我们开发了一种利用鼻窦计算机断层扫描（CT）图像的深度学习方法，以测试是否能自动检测出毛霉菌病，这可能有助于或加快传统上依赖活检的诊断过程。
在这项回顾性研究中，分析了794张CT图像（来自毛霉菌病患者、鼻息肉患者或正常检查结果患者）。对图像进行了调整大小和增强处理以用于训练。对两个迁移学习模型（ResNet50和ConvNeXt Small）进行了微调，以将图像分为三类。我们采用70/30的训练-测试分割（进行五折交叉验证），并使用准确率、精确率、召回率、F1分数和混淆矩阵来评估性能。
ConvNeXt Small模型在测试集上达到了100%的准确率（所有类别精确率/召回率/F1分数 = 1.00），而ResNet50模型达到了99.16%的准确率（精确率≈0.99，召回率≈0.99）。交叉验证产生了一致的结果（ConvNeXt在各折中的准确率约为99%），表明没有过拟合现象。一项消融研究证实了迁移学习的益处，因为从零开始训练ConvNeXt会导致较低的准确率（约85%）。结论：我们的研究结果表明，深度学习模型可以从CT扫描中准确且无创地检测出毛霉菌病，有可能标记出疑似病例以便及时治疗。这些模型可以作为快速筛查工具来补充标准诊断方法（组织病理学），不过我们强调它们只是活检的辅助手段而非替代品。未来的工作应在外部数据集上验证这些模型，并研究将它们整合到临床工作流程中以便对毛霉菌病进行更早干预。
