Wavelet Spectrum in a Multi‐Channel Network May Reduce Biopsy Rates on Diagnosis of Breast Tumors With BI ‐ RADS Category 4a or Higher

OBJECTIVES: BI-RADS 4a+ breast tumors have a high rate of unnecessary biopsies due to ambiguous diagnostic features. This study aims to develop and evaluate a computer-aided diagnosis method integrating wavelet spectrum analysis with deep learning for benign-malignant classification of these tumors, so as to reduce unnecessary invasive biopsies. METHODS: A retrospective study was conducted on 390 patients with breast lesions categorized as BI-RADS 4a or higher. We proposed a computerized diagnostic method that combines wavelet spectrum analysis with deep learning. A 3-channel ultrasound image was constructed by integrating B-mode, high-frequency (13 MHz), and low-frequency (7 MHz) wavelet components. This multi-channel input was processed by a dual-channel deep learning network for tumor classification. A strict patient-specific biopsy-sparing criterion was also established and evaluated. RESULTS: The integration of wavelet spectral features improved tumor classification performance by an average of 2% compared to using B-mode images alone. The proposed model achieved an overall accuracy of 84.52%. Under a patient-specific evaluation framework more aligned with clinical practice, the model demonstrated a biopsy-sparing rate of 26.01% while maintaining a low missed diagnosis rate of 1.786% at the most stringent threshold. The biopsy-sparing to missed diagnosis rate ratio remained high across thresholds, confirming the robustness of the patient-based assessment. CONCLUSION: The integration of wavelet spectrum analysis into a multi-channel deep learning network effectively improves the classification of BI-RADS 4a or higher breast tumors. This method demonstrates significant potential as a non-invasive adjunct tool to enhance diagnostic precision and reduce unnecessary biopsies in clinical practice.