Related Experiment Video
Updated: Aug 14, 2026
02:09
Multi-modal Pulmonary Imaging: Using Complementary Information from CT and Hyperpolarized 129Xe MRI to Evaluate Lung Structure-Function
Published on: April 12, 2024
CALM-CXR: a calibration-aware lung-masked workflow for hierarchical chest X-ray classification
Suresh Kumar Samarla1, D N S B Kavitha1, Solleti Phanikumar2
1Department of Computer Science and Engineering, Sagi Rama Krishnam Raju Engineering College, Bhimavaram, Andhra Pradesh, India.
Methodsx
|August 13, 2026
Summary
The CALM-CXR protocol offers an auditable method for evaluating chest X-ray classification by separating normal-abnormal screening from bacterial-viral subtyping. This approach improves performance by isolating influences like anatomical guidance and staged decisions.
Area of Science:
- Medical Imaging Analysis
- Machine Learning in Healthcare
- Radiology Informatics
Background:
- Chest X-ray classification models often lack transparency in how specific components influence performance.
- Existing studies frequently report overall metrics without dissecting the impact of anatomical guidance, staged decisions, or probability calibration.
- Error propagation in complex classification pipelines remains a challenge for accurate assessment.
Purpose of the Study:
- To introduce the CALM-CXR (Classification Auditing and Logging Method for Chest X-rays) protocol for reproducible and auditable evaluation of chest X-ray classification.
- To operationally couple established components like lung-mask conditioning, staged decision-making, and probability calibration in a controlled manner.
- To enable attribution of errors to specific stages (screening vs. subtyping) within the classification pipeline.
Main Methods:
- Developed CALM-CXR, a protocol coupling lung-mask-conditioned regional expert modeling with separate normal-abnormal and bacterial-viral classification stages.
- Implemented stage-specific temperature scaling, calibration-partition threshold selection, and coherent three-class probability composition.
- Conducted evaluations on 4,840 images using filename-derived group-disjoint partitions and performed component ablations to assess individual contributions.
Main Results:
- The full CALM-CXR protocol achieved accuracy of 0.7920 (SD 0.0248) and macro-F1 of 0.7999 (SD 0.0165) across multiple seeds.
- Ablation studies indicated that removing lung-mask conditioning resulted in the most significant decrease in accuracy and macro-F1.
- Matched flat ConvNeXt-Tiny baselines showed comparable performance, suggesting the hierarchical organization's superiority was not consistently established.
Conclusions:
- CALM-CXR serves as a valuable auditable evaluation protocol for chest X-ray classification, rather than a novel primitive or a clinically validated system.
- The protocol successfully separates and evaluates screening and subtyping stages, retaining stage-specific probabilities and error attribution.
- Independent external validation is necessary to confirm the clinical utility and generalizability of the CALM-CXR protocol.
