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

Presurgical Upgrade Prediction of DCIS to Invasive Ductal Carcinoma Using Time-dependent Deep Learning Models with DCE MRI

https://pubmed.ncbi.nlm.nih.gov/38900042/

1.0

John D. Mayfield, email: ude.dravrah.hgm@dleifyamdj

National Cancer Institute/NIH R01CA249016; Moffitt Cancer Center Support Grant P30-CA076292; institutional support from H. Lee Moffitt Cancer Center Quantitative Imaging Core.

IRB-approved, HIPAA compliant; waiver of informed consent.

Deep learning classification using CNN and time-dependent CNN-RNN. Backbones: VGG16 and ResNet50 (with/without ImageNet pretraining). Best-performing architecture: VGG16 feature extractor + LSTM. Optimizer: Adam; learning rate ~1e-6 (tuned to 1e-5); early stopping (patience 5 after 30 epochs); micro-batch size 4; LSTM internal and recurrent dropout 0.5. Class imbalance handled with ADASYN; image augmentations via Keras (RandomContrast, RandomFlip, RandomSkew, RandomBrightness).

Authors state final model and weights available for 2 years upon publication via GitHub (account: radres2019); contact corresponding author for access/details.

Assist in preoperative prediction of DCIS lesions at risk of upgrade to invasive malignancy, potentially informing surgical planning (e.g., sentinel lymph node biopsy) and enabling de-escalation for low-risk lesions.

Clinical decision support during preoperative planning for patients with biopsy-proven DCIS undergoing breast MRI.

Automated image-based classification; no lesion segmentation prerequisite (uses cropped breast slice). Intended to support, not replace, clinician decision-making.

Women with core-biopsy–proven ductal carcinoma in situ undergoing preoperative dynamic contrast-enhanced breast MRI to estimate risk of upgrade to invasive disease.

Preoperative DCE breast MRI post-contrast phases: four axial fat-suppressed T1-weighted 3D GRE postgadolinium images. For static CNN: single axial slice (largest lesion area) per phase, cropped to affected breast, normalized 0–255, resized to 224×224; multiphase images as channels. For time-dependent CNN-LSTM: four sequential postcontrast phases combined into a 4-frame 1 fps video (.avi).

Identify DCIS lesion, select axial slice with largest lesion area, crop to affected breast excluding mediastinum/contralateral breast/markers; normalize pixel values 0–255; resize to 224×224. Use ADASYN for minority class oversampling; apply Keras augmentations. Train with stratified 10-fold CV (80:10:10 split with randomized 10% holdout), Adam optimizer (LR tuned around 1e-6 to 1e-5), early stopping after 30 epochs (patience 5).

Single-institution retrospective cohort; small sample size (n=154; 25 upgrades); no external independent test set; majority White population; only cases with MRI correlate included; potential bias/generalizability limits; conversion to .avi for temporal modeling may reduce dynamic range; requires manual selection/cropping of slice with largest lesion area.

Use time-dependent, sequential multiphase DCE MRI inputs (all four postcontrast phases) with a VGG16+LSTM architecture; models without segmentation but with careful cropping performed best in this study.

Stratified 10-fold cross-validation with randomized 10% holdout set; hyperparameter tuning described; 10×10 CV t-tests for significance; occlusion maps used for model attention; data handling and anti-leakage measures reported.