https://atlas.rsna.org/schemas/2025-11/model.json

Detection Transformer (DETR) for detection, localization, and characterization of focal liver lesions on abdominal ultrasound

https://gitlab.inria.fr/hdadoun/focal-liver-lesions-us

1.0

rf.airni@nuodad.dnih

Supported in part by the French government through the 3IA Côte d’Azur Investments in the Future project managed by ANR (ANR-19-P3IA-0002).

Institutional review board–approved retrospective multicenter study (IRB00011591); informed consent waived.

Detection Transformer (DETR) vision transformer–based end-to-end object detection network fine-tuned for ultrasound FLL detection, localization, and characterization.

All code used for this study: https://gitlab.inria.fr/hdadoun/focal-liver-lesions-us

Assists in screening by detecting and localizing focal liver lesions and characterizing them as benign or malignant on abdominal ultrasound images; performance matched or exceeded experts in several tasks, potentially aiding nonexpert caregivers.

Patients’ triage / screening decision support during ultrasound interpretation.

Class-specific probability thresholds selected to maximize F1 score on validation; IOU threshold of 0.3 used to assign true positives for localization/classification.

Automated detection, localization, and classification; supports clinician decision-making.

Image-based aid for detecting, localizing, and characterizing focal liver lesions on noncontrast B-mode abdominal ultrasound images of adult patients (≥18 years) in hospital settings.

Preprocessed B-mode abdominal ultrasound images (JPEG converted from DICOM; US fan retained; text/graphics and biometric overlays removed).

Apply to noncontrast B-mode abdominal US images after de-identification and fan/overlay removal as described. Use class-specific probability thresholds (chosen on validation) and IOU > 0.3 to consider matches. Intended for images similar to those from the two participating hospitals and vendors in the study.

Retrospective study with small external test set (48 patients, 155 images). Reference standard for localization required unanimous adjudication, which may bias dataset and limit scalability. Poor-quality or uninterpretable images were excluded; performance on such images is unknown. Networks and experts were blinded to clinical context; real-world performance with clinical data may differ. Operator dependence of ultrasound may affect generalizability across sites/operators/vendors. Lower performance for cyst detection, likely due to class imbalance and confusion with vascular structures.

Research-use decision support to assist nonexpert and expert users in screening for and assessing focal liver lesions on abdominal US images.

Training details reported: 100 epochs; SGD with Nesterov momentum; learning rate 0.0048; batch size 4; momentum 0.9; weight decay 0.0001; warm-up first epoch; LR decay at epochs 75 and 90; data augmentation via learned bounding-box policies. Code repository provided.