Breast cancer classification based on the integration of diagnostic algorithms for calcifications and masses using a mixture of experts

Abstract

Purpose To investigate the effectiveness of an integrated deep-learning (DL) algorithm, the Mixture of Radiological Findings Specific Experts (MoRFSE), in breast cancer classification by imitating the diagnostic decision-making process of radiologists. Methods A total of 2,764 mammographic images (1,462 breast cancer, 248 benign lesions, and 1,054 normal breast tissue) from the TOMPEI-CMMD were used. The MoRFSE comprises three DL models: a gate network for categorization (gNet) and two classification expert networks (cExp and mExp) specialized in capturing the distinct characteristics of calcifications and masses, respectively. This structure imitates radiologists’ comprehensive diagnostic process by applying distinct algorithms for different lesion types. The classification performance of MoRFSE was compared with the Conventional ResNet18 model. 5-fold cross-validation was used, and performance was assessed using the area under the receiver operating characteristic curve (AUC). DeLong’s test was performed to evaluate statistical significance. Results MoRFSE achieved a significantly higher AUC (0.9616) compared to Conventional ResNet18 (0.9577, p = 0.001348). Conclusion By integrating specialized algorithms for calcifications and masses, the MoRFSE model effectively emulates radiologists’ diagnostic process. These findings suggest that MoRFSE has the potential to improve the accuracy of breast cancer diagnosis based on mammograms.