An Attention-Enhanced Deep Learning Model for Robust Identification of Atypical Mitotic Figures

Atypical mitotic figures (AMF) are valuable biomarkers for assessing tumour aggressiveness, but accurately detecting them is difficult because they appear infrequently and often differ only subtly from normal mitoses in histopathology images. To address this challenge, this work introduces an attentionaugmented deep learning framework built on a DenseNet-121 backbone and trained using the curated MIDOG Mitotic Figures Binary Classification Dataset. Several baseline CNNs and attention modules - CBAM, CCA, ECA, PSA, and Squeeze-and-Excitation (SE) - were evaluated, with SE consistently delivering the best performance when applied before global average pooling. A compact SE1D non-local attention block was also inserted at various depths of DenseNet-121 to determine the most effective placement, with Position 3 providing the highest gains. The final model, combining a global SE module with SE1D at Position 3, reached an accuracy of 89.17%, a precision of 89.91%, and an F1score of 89.08%, outperforming all other configurations. These outcomes underscore importance of strategically placed channelattention mechanisms for classification of mitotic subtypes.