A newly designed deep learning architecture evaluating three-dimensional chest computed tomography scans distinguished active from inactive pulmonary tuberculosis with high accuracy, according to retrospective study findings published in Cluster Computing. The computational model, designated 3D MSGC-ResNet50, was constructed to assist clinicians in categorizing stages of pulmonary tuberculosis from volumetric scan data, addressing key needs in timely disease management and transmission prevention.
To build and evaluate the classification framework, researchers examined a retrospective imaging dataset collected from the Public Health Treatment Center of Guiyang City in Guizhou Province. The investigative cohort contained one chest computed tomography volume from each of 500 individual patients. The study population was evenly distributed, comprising 250 individuals diagnosed with active pulmonary tuberculosis and 250 with inactive disease.
Prior to classification, the investigators implemented a dedicated 3D ResUNet model to segment the lung parenchyma across each patient scan volume.
The core classification pipeline was constructed around the 3D ResNet50 neural network architecture, enhanced through integrated multi-scale attention and grouped convolution techniques. Termed MSGC, the system combines a multi-scale attention module with grouped convolution featuring channel shuffle operations. These architectural modifications were engineered to improve the network’s capacity to capture multi-scale lesion characteristics and concentrate analytical focus on critical pathological areas.
During performance evaluations, the proposed 3D MSGC-ResNet50 model demonstrated strong diagnostic capability across standard statistical metrics. The system achieved a precision of 94.87 percent, an overall recall of 91.60 percent, and an F1-score of 93.19 percent. Additionally, the classification network produced an area under the receiver operating characteristic curve of 96.68 percent.
These performance outcomes represented measurable gains relative to the standard baseline 3D ResNet50 network without the multi-scale attention and grouped convolution enhancements. The modified architecture exceeded the baseline model by 3.90 percentage points in precision, 2.40 percentage points in recall, 3.21 percentage points in F1-score, and 4.71 percentage points in area under the curve.
The investigators also assessed the model’s diagnostic reliability across a separate external test set comprising 40 patient cases obtained from an independent second hospital. When applied to this external testing cohort, the 3D MSGC-ResNet50 system attained an overall classification accuracy of 90.00 percent, correctly categorizing 36 patient cases while recording misclassifications in four instances.
The researchers concluded that these preliminary external testing results offer initial empirical evidence supporting the network’s cross-institutional classification performance. The authors noted that the architecture demonstrates practical potential for assisting clinicians in differentiating active and inactive pulmonary tuberculosis, while highlighting the preliminary nature of the external hospital validation.
Subject of Research: Deep learning classification of active and inactive pulmonary tuberculosis on chest CT
Article Title: 3D MSGC-ResNet50: classification of active and inactive pulmonary tuberculosis on chest CT
Article References: Shi, W., Xiao, L., He, Z., Wang, R., Zhang, M., & Huang, Z. (2027). 3D MSGC-ResNet50: classification of active and inactive pulmonary tuberculosis on chest CT. Cluster Computing, 30(1), Article 59. https://doi.org/10.1007/s10586-026-06827-x
Image Credits: AI Generated
DOI: 10.1007/s10586-026-06827-x
Keywords: pulmonary tuberculosis, active tuberculosis, inactive tuberculosis, chest computed tomography, 3D MSGC-ResNet50, 3D ResUNet, lung parenchyma segmentation, multi-scale attention module, grouped convolution, channel shuffle, deep learning classification, receiver operating characteristic, external validation, thoracic imaging, lesion feature extraction, Public Health Treatment Center of Guiyang City, Cluster Computing
News Source: Denise Maddox. (November 21, 2026). Enhanced 3D ResNet Model Classifies Active Pulmonary Tuberculosis on Chest CT. Scienmag.



