The likelihood of developing breast cancer (BC) in women is approximately 12. 24%, the highest among all cancers, and clinicians are adopting AI-based technologies to improve early diagnosis and reduce the mortality rate. The present study focuses on histopathological image analysis, particularly the segmentation of Gigapixel Whole Slide Images (WSIs). However, designing a clinically deployable model requires careful consideration to achieve high segmentation accuracy, fast inference speed, and a reduced model size. To address these requirements, this study introduces LightMuSeg, a novel WSI segmentation framework. LightMuSeg utilizes a pre-trained MobileNet-V2 encoder for feature extraction, thereby reducing model size and achieving faster inference speeds. Additionally, the network integrates two specialized modules: the Multi-Scale Feature Fusion Module (MSFFM), which captures multi-scale feature representations to mitigate scale variance across tissue structures, and the Feature Refinement and Fusion Module (FRFM), which enhances salient feature representation to handle segmentation issues related to object camouflage properties. LightMuSeg was trained using the BCSS dataset for binary segmentation and the BCSS-WSSS dataset for weakly supervised multiclass segmentation, both of which are publicly available. In our experimental study for binary segmentation, the LightMuSeg-V2 variant achieved the highest mean Dice (76. 51%) and IoU (66. 43%) outperforming other models. When evaluated on the BCSS-WSSS dataset for multi-class segmentation, LightMuSeg achieved state-of-the-art performance in the Lymphocyticᵢnfiltrate and Necrosisₒrdebris classes, with IoU scores of 77. 2% and 93. 89%, respectively. Furthermore, a faster inference speed (44. 51 mTps), better tumor segmentation accuracy, and a small model size (7. 05 million parameters) make the proposed model suitable for deployment in real-world clinical scenarios. • Multi-scale feature extraction and fusion are proposed for accurate yet efficient segmentation. • The proposed lightweight model approach achieved the highest inference speed. • At the same time, it achieves the best or second score in both binary and multi-class segmentation.
Muhammad et al. (Wed,) studied this question.