Hybrid texture-deep feature fusion for mammogram classification: a patient-level, calibrated evaluation

Abstrak

We propose a lightweight computer-aided diagnosis (CAD) framework that fuses four sub-band discrete wavelet transform gray-level co-occurrence matrix (DWT–GLCM) texture features with fine-tuned ResNet-50 embeddings under a strict, patient-level, leak-free evaluation protocol. Experiments were conducted on two public datasets: mammographic image analysis society (MIAS) (normal vs. abnormal) and curated breast imaging subset of the digital database for screening mammography (CBIS-DDSM) (benign vs. malignant). Five-fold cross-validation (CV) was confined to the training portion, operating thresholds were fixed on the validation split to target high recall, and the held-out test set was evaluated once. Performance was assessed using accuracy, F1-score, receiver operating characteristic (ROC)-area under the curve (AUC) with bootstrap 95% confidence intervals (CI), precision-recall (PR)-AUC, and calibration metrics (Brier score, expected calibration error). The proposed fusion model achieved ROC-AUC on MIAS (0.992) and strong performance on CBIS-DDSM (0.896), with consistent PR characteristics. Calibration analysis indicated reliable probability estimates and clinically interpretable decisions at a 95% sensitivity operating point. Ablation experiments revealed substantial gains over texture-only baselines and parity with convolutional neural network (CNN)-only models, highlighting fusion as a simple yet well-calibrated alternative for screening-oriented workflows. This study underscores the necessity of patient-level evaluation, explicit operating-point selection, and calibration reporting to ensure clinically meaningful CAD performance in mammography.