MMFM Breast-Imaging Baselines — Single-Modality ResNet-18 Checkpoints
ResNet-18 binary (benign vs. malignant) classification checkpoints, trained
per-dataset on eight public breast-imaging sources spanning ultrasound,
mammography, and DCE-MRI. Code: MMFM repo.
Full write-up, methodology, and comparison to each source paper's own
numbers: docs/latex/main.tex in the repo.
These are single-modality baselines, not the 3-branch fusion model.
Each checkpoint is models.SingleBackboneClassifier (one ResNet-18 backbone,
ImageNet-pretrained, first conv adapted for non-RGB inputs where applicable)
— see models/backbone.py / training/train.py in the repo for the
loading code.
Files
| File | Dataset | Modality | Evaluation | Test accuracy | Notes |
|---|---|---|---|---|---|
busbra_resnet18_fold1.pth |
BUS-BRA | Ultrasound | 1 of 5 official CV folds | 0.828 ± 0.023 (5-fold mean) | Paper-matched recipe; see Table 2 in the report |
busi_resnet18_fold1.pth |
BUSI | Ultrasound | 1 of 5 stratified CV folds | 0.884 ± 0.038 (5-fold mean) | Full 780-image release |
busc_resnet18_fold1.pth |
BUSC | Ultrasound | 1 of 5 stratified CV folds | 0.992 ± 0.010 (5-fold mean) | Near-perfect result flagged as unverified — Grad-CAM analysis in the report suggests possible shortcut learning (no ROI crop in this dataset's preprocessing); do not treat as a working classifier without further checks |
breast_lesions_usg_resnet18_fold1.pth |
BrEaST-Lesions USG | Ultrasound | 1 of 5 case-level CV folds | 0.726 ± 0.042 (5-fold mean) | High fold-to-fold variance, ~51 test cases/fold |
mias_resnet18_fold1.pth |
mini-MIAS | Mammography | 1 of 5 patient-level CV folds | 0.599 ± 0.096 (5-fold mean) | Weak result; misses ~51% of malignant cases (see per-class bias table in the report) |
breastdm_resnet18.pth |
BreastDM | DCE-MRI (9-channel img9Se) |
Official train/val split | 0.880 | Within 0.002 accuracy of the paper's own LG-CAFN fusion model; specificity is weak (0.542) |
cdd_cesm_resnet18.pth |
CDD-CESM | Mammography (contrast-enhanced) | Single patient-level split, smoke test | 0.614 | Not a tuned reproduction — no augmentation search or class weighting yet |
cmmd_resnet18.pth |
CMMD | Mammography | Single patient-level split, smoke test | 0.758 | Class-weighted CE; not a tuned reproduction — no CV, no multi-view aggregation |
For the 5-fold datasets, this is one fold's checkpoint, not an ensemble or
the averaged model — reported accuracy is the 5-fold mean from the full
report for context, not this specific file's number. See the repo's
docs/results_report.md and docs/per_class_results.md for per-fold and
per-class (sensitivity/specificity/FP-vs-FN bias) detail.
Known limitations (see the full report for detail)
- BUSC's near-perfect accuracy is flagged, not confirmed — Grad-CAM shows diffuse, non-lesion-localized activation, consistent with the dataset's lack of an ROI crop step.
- mini-MIAS, CDD-CESM, and CMMD are weak/preliminary results, not submission-ready baselines.
- Most datasets use image- or row-level splits, not patient-level, where the source release doesn't ship a patient identifier (BUSI, BUSC) — possible near-duplicate leakage across folds.
Source data attribution
- BUS-BRA: Gómez-Flores et al., 2024, Medical Physics, 10.1002/mp.16812
- BUSI: Al-Dhabyani et al., 2020, Data in Brief, 10.1016/j.dib.2019.104863
- BUSC: Mendeley Data
- BrEaST-Lesions USG: Pawlowska et al., 2024, Scientific Data, 10.1038/s41597-024-02984-z, CC BY 4.0
- mini-MIAS: Suckling et al., 1994
- BreastDM: Zhao et al., 2023, Computers in Biology and Medicine, 10.1016/j.compbiomed.2023.107255
- CDD-CESM: Khaled et al., 2022, Scientific Data, 10.1038/s41597-022-01238-0, CC BY 4.0, via TCIA
- CMMD: Cui et al., 2021, The Cancer Imaging Archive, 10.7937/tcia.eqde-4b16, CC BY 4.0, via TCIA
Intended use
Research and educational use for breast-imaging CAD baselines and reproduction studies. Not validated for clinical use. Several of these checkpoints (BUSC, mini-MIAS, CDD-CESM, CMMD) have known weaknesses documented above and in the linked report — read those sections before using any of these checkpoints for anything beyond a baseline comparison.