APSR2026-572

 
Late-Breaking
Submitted
Oral Presentation
AO35-001
 
 
https://storage.unitedwebnetwork.com/files/1285/1215921-46903-APSR2026_440105.pptx
Format: PPTX Only
 
Format: PDF Only
14. Respiratory Structure and Function
3

No more than 15 authors can be listed.
Please ensure the authors are listed in the right order.
If there are multiple affiliations, please use the "+" button next to "Select Country/Region" to add additional affiliation fields.

Chia-Yi Chou * Taipei Medical University Graduate Institute of Biomedical Informatics Taipei Taiwan
Pai-Chien Chou - Taipei Medical University Hospital Division of Pulmonary Medicine, Department of Internal Medicine Taipei Taiwan Taipei Medical University Division of Thoracic Medicine, Department of Internal Medicine Taipei Taiwan
Hung-Wen Chiu - Taipei Medical University Graduate Institute of Biomedical Informatics Taipei Taiwan
 
 
 
 
 
 
 
 
 
 
 
 
Presenting Author Details
Chia-Yi
Chou
m610113002@tmu.edu.tw
+886900637632
Under 30
Taiwan
Original Research
Deep learning classification of spirometry flow-volume loop morphology using a ResNet-18 convolutional neural network
Maximum 25 words
Abstract Content
Visual interpretation of spirometry flow-volume loops is subjective, while traditional computational methods require manual feature extraction. This study evaluated the feasibility of a deep learning model for automated, image-based classification of flow-volume loop morphology to identify respiratory abnormalities.
A ResNet-18 convolutional neural network evaluated 154 spirometry images across six morphological classes. To address class imbalance, oversampling and weighted cross-entropy loss were applied. Model performance was assessed using stratified 3-fold cross-validation, evaluating metrics including area under the curve (AUC), accuracy, precision, recall, F1-score, and confusion matrices.
The model achieved class-specific AUCs ranging from 0.677 to 0.906. It demonstrated strong discrimination for obstructive (Class B, AUC: 0.906) and restrictive-like (Class D, AUC: 0.883) patterns. Overall accuracy was 0.56. Confusion matrices revealed significant overlap between normal (Class 0) and variable extrathoracic obstruction (Class A) cases, with 21 normal loops misclassified as Class A and 15 Class A loops misclassified as normal. Initial Grad-CAM heatmaps showed the model over-focused on the peak expiratory flow area, neglecting inspiratory features. Post-intervention Grad-CAM confirmed attention successfully shifted toward the descending limb and inspiratory phases, better matching clinical practice.
A ResNet-18 model demonstrates feasibility for automated classification of spirometric loops directly from images. Misclassifications frequently occurred between normal and variable extrathoracic patterns. This likely happened because distinguishing a subtle inspiratory plateau from normal variance is morphologically challenging, especially when the model initially biased its attention toward expiratory flow. Despite limitations from dataset size and class imbalance, these findings highlight the potential of deep learning to assist in automated pulmonary function test interpretation. Future studies integrating numerical spirometry data with image analysis are warranted to further enhance classification accuracy and overcome morphological ambiguities.
 
 
 
 
 
No
No
 
JPG or JPEG or PNG format ONLY
Spirometry, Flow-volume loop, Deep learning
maximum 3 keywords and no abbreviations, separated by commas
The authors declare no conflict of interest.
279