Spiking neural transformer based confidentiality preserving framework for breast cancer detection in histopathology images

Breast cancer diagnosis by medical imaging requires accuracy and high sensitivity to patient information. Even though artificial intelligence has made major progress in the field of histopathological image analysis, there are still limitations such as noise in the image, structural variability, and class imbalance, low interpretability of deep learning models, high computational complexity, and privacy risks. To overcome these limitations, this paper suggests a Confidentiality-Preserving Spiking Neural Transformer for Breast Cancer Detection (CP-SNT) to secure and robustly detect breast cancer. In initial phase Robust Double-Weighted Guided Image Filtration (RDWGIF) is applied to improve the quality of histopathological images and reduce noise, but not fine tissue borders. This is followed by the Discrete Quaternion Quadratic Phase Fourier Transform (DQQPFT) that obtains discriminative frequency and phase-domain features, allowing the representation of complex structural and textual patterns. The central classification layer is a Probabilistic Spiking Neural Network (PSNN) that is integrated with a Causality-Aware Transformer Network (CATN) spike-based uncertainty modeling and attention-directed learning, respectively. To maintain the privacy of data in training and inference, PrivTuner using Homomorphic Encryption allows the safe computation of medical data encrypted with the homomorphic encryption to avoid the disclosure of sensitive patient data. The proposed CP-SNT framework is evaluated on the publicly available BreakHis and Breast Histopathology Images (BHI) datasets of 16,116 images to validate classification performance across multi-scale and patch-level histopathological samples. Comprehensive experimental validation on the BreakHis and BHI datasets demonstrates that the proposed CP-SNT model achieves accuracies of 98.60% and 98.37%, respectively. The framework further records precision values above 98.1% and F1-scores near 98.2%, reflecting its strong capability to minimize both false positives and false negatives. Such outstanding performance highlights the effectiveness of the hybrid spiking-transformer architecture combined with encrypted-domain computation for secure and high-confidence breast cancer diagnosis.