Answers are grounded in this paper and cite their sources.
Background Current clinical pathogen diagnosis mainly relies on high-throughput sequencing and conventional culture, both of which have limitations in time and cost. This study developed a convolutional neural network-based Mix-inResNet model to support CT-based preclassification of pulmonary infection categories. Purpose This study was designed to develop and internally evaluate Mix-inResNet using CT images from microbiologically defined pulmonary infection cases. Methods A total of 273 patients, including patients in the four pulmonary-infection categories and separate normal-CT controls, were included. Labels for the infection categories were supported by metagenomic next-generation sequencing (mNGS) and conventional culture. The image-level dataset contained 2,858 original CT slice images derived from these patients. Mix-inResNet was compared descriptively with ResNet101, Inception V3, VGG16, clinicians, and imaging specialists in the internal test set. Results Mix-inResNet achieved precision, accuracy, F1 score, and Gwet’s kappa values of 94.61%, 94.41%, 94.22%, and 93.09%, respectively. For bacterial, fungal, viral, and tuberculous pneumonia, as well as normal chest CT, the sensitivities were 88.14%, 92.11%, 91.94%, 98.00%, and 100%, respectively, and the specificities were 98.68%, 100.00%, 97.77%, 97.46%, and 99.04%, respectively. The internal test-set accuracy was 94.41% with a Wilson 95% confidence interval of 91.11%–96.53%. At the patient level, 27 of 27 test patients were correctly classified (Wilson 95% CI, 87.54%–100.00%); because this estimate rests on only 27 patients, it cannot support patient-level generalization. In descriptive comparisons, Mix-inResNet showed high point estimates for several performance metrics and AUC values; these internal comparisons do not establish general superiority over other models or superiority, non-inferiority or equivalence relative to clinician readers. In an exploratory validation-set ablation of retrained variants, the full model had a validation accuracy of 93.71%; because the validation partition had also been used for hyperparameter tuning, module contributions were not regarded as confirmed. A real-slice replacement sensitivity analysis showed that accuracy fell from 94.41% to 56.99% and calibration worsened when 75% of the selected infection slices were replaced by previously unselected slices, indicating substantial spectrum bias that was quantified but not mitigated. The expanded reader comparison was classified as a pilot feasibility exercise and was not used to infer superiority or equivalence of the model relative to clinical readers. Grad-CAM was used to visualize image regions contributing to model predictions. Conclusions Mix-inResNet showed promising internal performance for CT-based preclassification of microbiologically defined pulmonary infection categories. The model does not directly identify microorganisms from CT images and should be regarded as an exploratory research model that requires larger prospective multicenter validation before any clinical application is considered. Because spectrum bias was not mitigated and patient-level evaluation included only 27 test patients, the reported performance is likely optimistic, clinical utility has not been established, and the model is not ready for clinical use.
No takes yet. Share an insight, caveat, or question.
Meng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: