Soran A. Pasha, Fadhil Salman Abed
Breast ultrasound imaging is a widely available, non-ionizing, and cost-effective modality for breast cancer screening; however, reliable multiclass classification remains challenging because of speckle noise, low contrast, irregular lesion boundaries, inter-scanner variability, and visual overlap among benign, malignant, and normal tissues.FiLMoS-Net integrates learnable Gabor orientation cues, block-wise DCT frequency representations, differentiable edge-morphology features, and sample-adaptive convex routing in a compact 1,288,966-parameter architecture.Evaluation used 780 independent BUSI images in a fixed stratified 544/118/118 train/validation/test split.All comparative runs used random initialization, the same optimization and validation-selection protocol, and five distinct training seeds (42-46).The seed-42 cold-start checkpoint achieved 91.53% accuracy, 91.80% macro-F1, 91.64% sensitivity, 95.11% specificity, 95.54% macro-AUROC, and 94.80% macro-AUPRC.Its benign, malignant, and normal F1 scores were 93.13%, 84.85%, and 97.44%, respectively.Five-run means, sample standard deviations, confidence intervals, and paired nonparametric comparisons were calculated from the archived per-sample predictions of the five independent runs.A separate warm-start sensitivity analysis achieved 93.22% accuracy and 93.71% macro-F1 and was excluded from the cold-start comparative statistics.