AI Reads the Diaphragm in Seconds on Routine Thoracic MRI, Multireader Study Finds

Researchers have developed and validated a fully automated artificial intelligence tool that outlines the diaphragm on routine thoracic magnetic resonance imaging, cutting a task that once took radiologists up to four and a half hours per scan down to as little as five seconds. The study, published in European Radiology Experimental, tested two deep learning models on both dynamic 2D-cine MRI and static 3D volumetric MRI of the chest, and found that the automated contours matched expert manual annotations closely enough that a panel of five blinded radiologists judged them clinically usable in the vast majority of cases.

The diaphragm is the principal muscle of breathing and a central determinant of how the chest and abdomen move together. It can be impaired in a wide range of conditions, including neuromuscular disorders, chronic respiratory diseases, side effects of cancer treatment, inflammatory myopathies, and abdominal disorders that alter the load on the muscle. Conventional ways of assessing it have drawbacks: fluoroscopy and computed tomography visualize the diaphragm directly but expose patients to ionizing radiation and offer mostly static or limited-angle views, while ultrasound is radiation-free but highly operator-dependent and often restricted to two-dimensional measurements.

MRI offers a noninvasive, radiation-free alternative with excellent soft-tissue contrast, capable of detailed two-dimensional cine and three-dimensional or four-dimensional analysis of diaphragmatic shape and motion. Recent applications in rare diseases such as thoracic insufficiency syndrome and Pompe disease have shown that MRI-derived diaphragm metrics can reveal subtle functional decline that conventional spirometry misses. Yet diaphragmatic MRI analysis has remained largely manual or semiautomated, with cumbersome phase selection that limits reproducibility, scalability, and translation into clinical practice. Fully automated volumetric segmentation of the diaphragm has been a key unmet need.

To address that gap, a single-center study led by Samia Boussouar, Alexiane Pasquier, Damien Vaurs, and colleagues, with Dima Rodriguez as senior author, trained two U-Net convolutional neural network models, one for 2D-cine images and one for 3D volumes, using expert manual contours as the reference standard. The cohort comprised adults without known or suspected diaphragmatic disorders who underwent clinically indicated thoracic MRI on a 1.5-Tesla scanner. In total, 59 individuals contributed 70 MRI acquisitions: 29 women and 30 men, with a mean age of 49.3 years. Forty analyzable 2D-cine acquisitions were retained after screening 246 consecutive clinical scans, and 30 analyzable 3D volumetric interpolated breath-hold examination acquisitions were retained after 5 of 35 were excluded for nondiagnostic image quality.

The reference annotations were performed by a thoracic radiologist with 13 years of experience, working slice by slice for the 2D dataset and with multiplanar refinement for the 3D dataset. Diaphragmatic pillars were excluded because of incomplete coverage. The models were trained with a hybrid loss combining Dice and binary focal cross-entropy, monitored to prevent overfitting, and assessed with five-fold participant-level cross-validation. All dataset splits were individual-disjoint, meaning no person contributed to more than one subset, and final performance was measured on an independent held-out test set using surface-based metrics chosen because the diaphragm is a thin, widely extended surface occupying only a small fraction of the image volume, which makes traditional volumetric overlap measures uninformative.

The quantitative results were strong on both fronts. The 2D model achieved a normalized surface Dice of 0.95 plus or minus 0.08 at a 3-voxel tolerance, with an average surface distance of 1.97 plus or minus 2.29 millimeters. The 3D model reached a normalized surface Dice of 0.92 plus or minus 0.06, with an average surface distance of 1.88 plus or minus 1.97 millimeters. The 95th percentile Hausdorff distance was 9.89 plus or minus 15.2 millimeters for 2D and 7.67 plus or minus 6.3 millimeters for 3D. For context, the researchers note that diaphragm excursion during breathing is typically on the order of several centimeters, so a boundary deviation of a few millimeters represents a small fraction of expected displacement.

Speed was equally striking. Manual annotation required about 10 minutes per 2D image and 4.5 hours per 3D volume, whereas automated inference took 5 seconds per 2D image and under 30 seconds per 3D volume on a single graphics processing unit, more than a hundredfold faster. Five-fold cross-validation showed consistent performance with low inter-fold variability, and the 3D model in particular was described as stable and relatively insensitive to image noise.

Beyond the numbers, five board-certified radiologists with a range of experience independently reviewed segmentations from data outside the training, validation, and testing sets, without involvement in study design or model development. They rated segmentation quality on a four-point scale and image quality on a five-point scale, with clinical acceptability requiring a quality score of at least 3 out of 4 and a checklist score of at least 3 out of 5 by majority consensus. In the main cohort, segmentations were judged clinically acceptable in 87.5 percent of 2D cases, 14 of 16, and in 100 percent of 3D cases, 9 of 9. Mean image-quality scores were 4.0 out of 5 for 2D and 4.2 out of 5 for 3D, and image quality correlated moderately to strongly with segmentation performance, with a correlation coefficient of 0.65 for 2D and 0.64 for 3D, the latter not reaching statistical significance.

The team also examined a small separate feasibility cohort of five patients with diaphragmatic dysfunction, assessed by two blinded thoracic radiologists using the same scale. Usability for 2D images was 75 percent for one reader and 70 percent for the other, with agreement on 17 of 20 images, and 80 percent for both readers on 3D volumes, with agreement on 8 of 10. The authors stress that this pathology cohort was included for feasibility assessment only and does not constitute clinical validation, and that primary model development was performed in subjects without known or suspected diaphragmatic dysfunction to establish a controlled baseline.

Typical failure modes were instructive. Discrepancies clustered in the costophrenic recesses, where reduced image quality limited contour conspicuity, suggesting limited visibility rather than systematic algorithmic bias. Occasional false positives appeared in upper slices, where shoulder structures with similar intensity and curvature were mislabeled as diaphragm, an error the authors suggest could be mitigated with rule-based postprocessing or anatomical continuity constraints. Compared with prior work, which relied mostly on indirect measures or pediatric cohorts, the researchers describe this as one of the first adult references for direct diaphragm segmentation on dynamic MRI. Limitations include the single-center setting, a predominantly healthy cohort, a single-radiologist reference standard, and internal testing, all of which may restrict generalizability, particularly to patients with complex disease and lower-quality imaging. The study should be considered a proof of concept, and the authors call for larger multicenter and disease-cohort validation before the approach can support routine, standardized diaphragm biomarkers in clinical workflows.

Subject of Research: Fully automated deep learning segmentation of the diaphragm on routine 2D-cine and 3D thoracic MRI, validated against expert manual contours and reviewed by multiple blinded radiologists.

Article Title: Fully automated diaphragm segmentation on routine 2D-cine and 3D thoracic MRI: multireader validation

Article References: Boussouar, S., Pasquier, A., Vaurs, D., Douho Kamta, A. D., Reitmann, A., Pasi, N., Donciu, V., Benchara, S., Wong, T., Belghiti, M., Gharmaoui, I., Dubuisson, R.-M., Nemeth, A., Lebon, V., Pellot-Barakat, C., Poirier-Quinot, M., Maître, X., Redheuil, A., & Rodriguez, D. (2026). Fully automated diaphragm segmentation on routine 2D-cine and 3D thoracic MRI: multireader validation. European Radiology Experimental, 10(1), Article 159. https://doi.org/10.1186/s41747-026-00805-z

Image Credits: AI Generated

DOI: 10.1186/s41747-026-00805-z

Keywords: diaphragm, MRI segmentation, deep learning, U-Net, 2D-cine MRI, 3D VIBE MRI, normalized surface Dice, radiologist evaluation, respiratory mechanics, medical imaging automation, Fully, automated