Lightweight AI network reads chest X-rays to tell bacterial from viral pneumonia

Researchers in China have developed a compact deep learning model that can distinguish bacterial pneumonia, viral pneumonia, and healthy chest X-rays with 91.4 percent accuracy while using a fraction of the computing power demanded by many competing artificial intelligence systems. The model, called FLIFuse, was described in a study published in the journal Frontiers in Cellular and Infection Microbiology, and its designers say it is intended for hospitals and clinics where high-end hardware and experienced radiologists are in short supply.

Pneumonia remains one of the deadliest lower respiratory tract infections worldwide, and telling its bacterial and viral forms apart matters enormously for treatment. The two types can look strikingly similar on a chest X-ray and in early clinical symptoms, yet they demand fundamentally different management: bacterial infections call for antibiotics, while viral cases do not respond to them. Misdiagnosis therefore risks both delayed treatment and unnecessary antibiotic use, which fuels antimicrobial resistance. Laboratory tests, meanwhile, suffer from long turnaround times, low positive rates, and sensitivity to sampling timing and pathogen load, leaving imaging interpretation heavily dependent on physician expertise.

To tackle this, a team from Huzhou Central Hospital and Huzhou Normal University, working with Jinhua Guangfu Oncology Hospital, built FLIFuse around a three-branch parallel architecture. Each branch plays a distinct role. A local feature extraction branch uses a module called ExtraDW, borrowed from MobileNetV4, which relies on depthwise separable convolutions to capture fine-grained details such as lesion edges, textures, ground-glass opacities, and small nodules. A global branch uses a slimmed-down module the authors call Conv2Next, inspired by the ConvNeXt architecture, to model broader semantic structures such as lung field texture and the overall distribution of lesions, helping separate diffuse interstitial changes from segmental consolidation.

The third component, the FLI Block, is where the fusion happens. Global features pass through a channel attention mechanism that dynamically recalibrates the weights of feature channels, emphasizing those most useful for distinguishing pneumonia types. Local features pass through a spatial attention module that highlights lesion-related regions and suppresses background noise. The two attention-processed streams are then integrated with the output of the previous fusion layer through a complementary balancing mechanism, and the combined features are refined by a convolutional residual block before a classifier produces the final decision. The authors report that this design mitigates feature redundancy and helps preserve the discriminative information that separates bacterial from viral pneumonia.

On the public ChestXRay2017 dataset, which contains 5,856 pediatric chest X-rays graded by professional physicians across the three categories, FLIFuse reached 91.4 percent accuracy with a recall of 72.7 percent, precision of 76.8 percent, and an F1 score of 74.4 percent. Those figures matched or exceeded larger architectures including ConvNeXt, HiFuse, ViT, and the Swin Transformer, which achieved 91.0, 91.4, 89.9, and 91.0 percent accuracy respectively. Crucially, FLIFuse did so with only 33.43 million parameters and 6.61 billion floating-point operations per inference, far below the computational budgets of HiFuse at 132.31 million parameters and 20.24 GFLOPs, ConvNeXt at 87.51 million parameters, and the Swin Transformer at 86.68 million parameters.

The resource savings extend to storage and memory. Under INT8 quantization, the model’s weight file weighs in at roughly 33.4 megabytes, about 74.7 percent smaller than the baseline model’s 132.3 megabytes. The authors estimate peak activation memory at approximately 5 to 10 megabytes for a single 224-by-224-pixel input at FP32 precision, a level they describe as low enough for deployment in computationally constrained environments such as primary care facilities.

Interpretability was assessed with Grad-CAM heatmaps, which visualize where the model focuses when making decisions. For correctly classified viral pneumonia cases, highlighted regions concentrated on areas of interstitial exudation and diffuse ground-glass opacities with a relatively spread-out activation pattern, reaching 98.31 percent confidence in one displayed example. Bacterial pneumonia cases showed more concentrated focal hotspots consistent with lobar or segmental consolidation, with one example reaching 95.93 percent confidence. Normal X-rays produced no concentrated high-response regions within the lung fields. The authors note that misclassified samples also yielded useful clues: in one false positive, the model responded strongly to normal lung markings, while in one false negative the heatmap showed visible activation in the lesion region that fell below the classification threshold, suggesting that threshold optimization or multi-scale features could improve stability on low-confidence cases.

Cross-domain testing provided evidence that the architecture transfers beyond its original task. On a COVID-19 chest X-ray database containing 1,200 COVID-19-positive, 1,341 normal, and 1,345 viral pneumonia images, FLIFuse achieved 88.1 percent accuracy, 78.8 percent recall, 83.3 percent precision, and an F1 score of 80.4 percent, outperforming ConvNeXt, EfficientNet, ViT, HiFuse, and MobileNetV4 on overall metrics. On the Nerthus endoscopic dataset, which assesses bowel preparation quality across 5,525 frames from 21 colonoscopy videos, the model reached 94.7 percent accuracy and an F1 score of 85.8 percent, trailing only ViT on accuracy by 0.2 percentage points while using far fewer parameters. The authors interpret this as evidence that the features the network learns are not narrowly tied to chest radiographs.

Statistical robustness was examined through five-fold cross-validation on all three datasets, with results compared against HiFuse, the baseline architecture from which FLIFuse draws its hierarchical fusion inspiration. On ChestXRay2017, FLIFuse posted a mean accuracy of 91.40 percent with a standard deviation of 0.38, compared with 90.18 plus or minus 1.13 for HiFuse, alongside an F1 score of 78.11 percent and specificity of 95.20 percent. On the COVID-19 dataset it reached 88.02 percent accuracy and an 80.20 percent F1 score, and on Nerthus 94.34 percent accuracy with an 84.21 percent F1 score, roughly 3.6 percentage points above HiFuse. The authors acknowledge that some confidence intervals overlap between the two models, but they report that the direction of improvement was consistent across folds.

Ablation experiments supported the three-branch design. A local-branch-only model achieved 89.0 percent accuracy on ChestXRay2017 and a global-branch-only model 89.4 percent, while the full architecture reached 90.5 percent in that configuration test, with gains in precision and F1 score as well. Replacing the ExtraDW module with ResNet, RepVGG, PConv, MobileOne, or GhostNetV3 blocks reduced performance, with the ResNet block dropping accuracy to 84.5 percent, and substituting the Conv2Next block with ViT, Swin Transformer, or ConvNeXt modules also lowered accuracy. The authors caution that the study rests on retrospective public datasets with relatively concentrated sample sources, that evaluation was largely single-center, and that prospective clinical validation and deployment engineering remain future work. Even so, they argue the results establish FLIFuse as a candidate for computer-aided pneumonia interpretation where accuracy and efficiency must both be respected.

Subject of Research: A lightweight deep learning model for distinguishing bacterial and viral pneumonia on chest X-rays

Article Title: FLIFuse: a lightweight parallel multi-branch network for pneumonia pathogen identification in chest x-rays

Article References: Yang, Y., Yao, Z., Zhang, Z., Yu, M., Yao, C., Jia, L., & Hua, F. (2026). FLIFuse: a lightweight parallel multi-branch network for pneumonia pathogen identification in chest x-rays. Frontiers in Cellular and Infection Microbiology, 16, 1938619. https://doi.org/10.3389/fcimb.2026.1938619

Image Credits: AI Generated

DOI: 10.3389/fcimb.2026.1938619

Keywords: pneumonia, chest X-ray, deep learning, FLIFuse, computer-aided diagnosis, bacterial pneumonia, viral pneumonia, lightweight neural network, medical imaging, Grad-CAM, COVID-19, cross-domain transfer