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

DeepContrast: EfficientNetB4-based detection of intravenous contrast enhancement on CT

https://github.com/AIM-Harvard/DeepContrast

1.0

ude.dravrah.icfd@nnaK_nimajneB

Supported by NIH (H.J.W.L.A.: U24CA194354, U01CA190234, U01CA209414, R35CA22052; B.H.K.: K08DE030216), European Union–European Research Council (H.J.W.L.A.: 866504), Radiological Society of North America (B.H.K.: RSCH2017), and the National Institute of General Medical Sciences (D.P.: T32-GM007753).

Conducted in accordance with the Declaration of Helsinki; local IRB approval obtained with waiver of consent due to use of public datasets and retrospective design.

Convolutional Neural Network; EfficientNetB4 backbone (also evaluated ResNet101V2, InceptionV3, simple CNN; transfer learning tested). Implemented in TensorFlow 2.0.

Open-source code and trained model available at https://github.com/AIM-Harvard/DeepContrast

Automates reliable detection of intravenous contrast enhancement on CT scans to support data curation, quality assurance, and potential clinical workflow stratification of contrast-enhanced vs non-contrast studies.

Data curation/quality assurance; workflow optimization; potential integration into radiology reporting workflows.

Probability threshold of 0.5 for image- and patient-level classification; patient-level score is the average of image-level probabilities within a scan.

Fully automated scan-to-prediction pipeline; no manual interaction required for inference.

Binary detection of presence or absence of intravenous contrast enhancement on head and neck and chest CT scans for research data curation and potential clinical workflow support in radiology departments.

DICOM CT scans (head and neck or chest); processed as axial 2D image sections; patient-level probability obtained by averaging image-level probabilities.

Preprocess CT scans per described pipeline (co-registration for HN, crop to region, extract 2D axial sections, convert to NumPy arrays). Use trained EfficientNetB4 model; classify images and average to patient-level. Default decision threshold 0.5. Authors recommend small local tests on institutional data prior to large-scale implementation.

Development and validation limited to head and neck and chest CT; datasets with single contrast phase and known cancer diagnoses. Potential uncaptured confounders; known error modes include faint contrast, artifacts, and dense vessels causing false positives/negatives. Users should locally validate before deployment at scale.

Use for automated curation and QA of CT datasets; conduct small, local validation on institutional scans prior to large-scale use; can aid stratification of contrast-enhanced vs non-contrast studies in workflow.

Code, trained model, and measured results for statistical replication are publicly available. Implemented in Python 3.8/TensorFlow 2.0; trained/evaluated on an NVIDIA Titan RTX GPU. External validation performed on independent cohorts.

Approximate runtime including preprocessing, data loading, and prediction: 2.1 hours for 1,315 HN scans (86,790 axial sections) and 1.1 hours for 664 chest scans (46,690 axial sections) on a Titan RTX GPU.