CBTI-COS++: A Cognitive Biological-Twin Imaging Network with Deviation-Contrastive Learning and Validation-Calibrated Decisions for Breast Ultrasound Classification

Abstract

Abstract—Breast ultrasound is a widely used, radiation-free modality for breast-cancer screening, but automated benign/malignant classification remains difficult because of small labelled datasets, severe class imbalance, operator-dependent image quality, and the absence of complementary modalities such as MRI, CT, or PET in routine practice. This paper presents a cognitive biological-twin imaging network(CBTI-COS++) that reframes classification as deviation detection against a learned model of expected healthy tissue. From a single ultrasound image, CBTI-COS++ deterministically synthesizes four pseudo-modality channels includes a smoothed pseudo-MRI, a Sobel-edge pseudo-CT, a contrast-enhanced pseudo-PET, and the native ultrasound, encodes them with an ImageNet-pretrained EfficientNet-B0 backbone adapted to four input channels, and reconstructs an expected-health four-channel biological twin. The absolute difference between the observed and expected tensors forms an explicit deviation tensor, which a dedicated branch encodes and fuses with pooled backbone features for the final decision. Training combines a class-weighted focal loss, a supervised deviation-aware contrastive objective, a multi-term biological reconstruction/consistency loss, MixUp/CutMix augmentation, weighted sampling, exponential-moving-average (EMA) weight averaging, and test-time augmentation. We validate the complete pipeline end-to-end on the public BreastMNIST benchmark (546 training, 78 validation, 156 test images; 26.9% malignant). Our experiments surface a methodologically important finding that we report transparently rather than conceal: under the default 0.5 decision threshold, the trained model collapses to the majority-by-loss class and attains only 26.9% test accuracy despite a strong test ROC-AUC of 0.836, a textbook manifestation of the mismatch that class-weighted focal loss and heavy augmentation can induce between a model's ranking quality and its argmax calibration on an imbalanced dataset. Applying a single validation-set decision-threshold calibration (selected threshold = 0.33) recovers 73.1% test accuracy, 74.1% balanced accuracy, and 0.744 weighted F1 at the same AUC, with no change to the learned weights. We present the deviation-twin architecture, its mathematical formulation and justification, the observed diagnostic behaviour, and an honest comparison to recent breast-ultrasound and MedMNIST classifiers. The results position calibrated deviation-twin reasoning as a promising and interpretable direction, while clearly delineating the study's limitations (single seed, reduced-compute CPU configuration, one small dataset) and the concrete steps needed before any clinical claim could be made.