MD-Mamba: A multi-scale dilated state-space network for breast cancer histopathology image classification.
L’essentiel
Automated breast histopathology classification requires models that can capture both local cellular morphology and broader tissue architecture while remaining computationally efficient. We developed MD-Mamba, a multi-scale dilated state-space network with dual-path spatial-channel attention for four-class breast histopathology image classification. The model integrates selective state space modeling for long-range tissue organization, multi-dilated convolutions for local morphology at complementary receptive fields, and attention modules for spatial and channel-wise feature refinement. MD-Mamba was evaluated on the ICIAR 2018 BACH dataset using a stratified 80:20 split and compared with CNN, transformer, and hybrid baselines under the same protocol. In this benchmark split, MD-Mamba achieved 0.9625 accuracy, 0.9654 macro-precision, 0.9625 macro-recall, 0.9617 macro-F1, 0.9956 AUC, and 0.9500 Cohen's kappa. Errors were concentrated at the Normal-Benign boundary. Ablation and sensitivity analyses supported the contribution of MDCM, spatial attention, and channel attention. MD-Mamba showed strong benchmark performance for BACH four-class breast histopathology classification. External whole-slide, multi-institutional, patient-independent, and molecularly annotated validation is required before claims about clinical translation or biomarker utility can be made.
Synthèse détaillée
Résumé original
Automated breast histopathology classification requires models that can capture both local cellular morphology and broader tissue architecture while remaining computationally efficient. We developed MD-Mamba, a multi-scale dilated state-space network with dual-path spatial-channel attention for four-class breast histopathology image classification. The model integrates selective state space modeling for long-range tissue organization, multi-dilated convolutions for local morphology at complementary receptive fields, and attention modules for spatial and channel-wise feature refinement. MD-Mamba was evaluated on the ICIAR 2018 BACH dataset using a stratified 80:20 split and compared with CNN, transformer, and hybrid baselines under the same protocol. In this benchmark split, MD-Mamba achieved 0.9625 accuracy, 0.9654 macro-precision, 0.9625 macro-recall, 0.9617 macro-F1, 0.9956 AUC, and 0.9500 Cohen's kappa. Errors were concentrated at the Normal-Benign boundary. Ablation and sensitivity analyses supported the contribution of MDCM, spatial attention, and channel attention. MD-Mamba showed strong benchmark performance for BACH four-class breast histopathology classification. External whole-slide, multi-institutional, patient-independent, and molecularly annotated validation is required before claims about clinical translation or biomarker utility can be made.