Citation

Hasan, Md Jahid and Ahmad, Wan Siti Halimatul Munirah Wan and Ahmad Fauzi, Mohammad Faizal and Abas, Fazly Salleh and Lee, Jenny Tung Hiong and Khor, See Yee and Looi, Lai Meng and Khalid, Ahmad S. and Tang, Tong Boon and Razak, Normy N. (2026) LACE-Net: A novel lightweight attention-based CNN with knowledge distillation for efficient scoring in histopathological images. Biomedical Signal Processing and Control, 126. p. 110873. ISSN 17468094

Text 11.pdf - Published Version Restricted to Repository staff only Download (6MB)

Abstract

Histopathological biomarker scoring is crucial for cancer diagnosis and treatment but remains difficult due to subjectivity, inter-observer variability, and inconsistent quantification of staining intensity and proportion. In ER-IHC, the Allred system (0–8) combines proportion and intensity but still yields variable results. HER2 assessment adds cost and complexity by requiring specific stained slides. Although deep learning has advanced automation, existing models remain computationally intensive, data-dependent, and hard to interpret, limiting clinical adoption. In this work, we propose LACE-Net, a lightweight and design-driven architecture centered on a novel attention-integrated convolutional block and its stage-wise deployment for efficient histopathological feature modeling. Unlike conventional methods, LACE-Net embeds attention directly within its building blocks and aligns attention type with feature hierarchy, enabling effective feature extraction and adaptive refinement under computational constraints. To improve deployment feasibility, a knowledge distillation framework compresses the model from a 64K-parameter teacher to an 18K-parameter student with minimal performance degradation. LACE-Net achieves state-of-the-art accuracy—95.76 ± 0.11% (F1: 0.8919 ± 0.0017) on ER-IHC, 97.15 ± 0.1118% (F1: 0.9737 ± 0.0012) on BCI H&E, 96.52 ± 0.1525% (F1: 0.9651 ± 0.0016) on BCI HER2, and 95.97 ± 0.0460% (F1: 0.9548 ± 0.0003) on CRC-100K, while the student model maintains competitive performance with significantly fewer parameters. Ablation studies confirm the effectiveness of the proposed block design and stage-wise attention strategy in capturing subtle morphological and staining variations. Furthermore, Grad-CAM++ and SHAP visualizations highlight diagnostically relevant nuclei and membrane regions, demonstrating the interpretability and clinical potential of LACE-Net for automated histopathology analysis.

Item Type:	Article
Uncontrolled Keywords:	Histopathological image analysis, breast cancer diagnosis
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1501-1820 Applied optics. Photonics
Divisions:	Faculty of Engineering and Technology (FET) Faculty of Artificial Intelligence & Engineering (FAIE)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	04 Aug 2026 06:32
Last Modified:	04 Aug 2026 06:32
URII:	http://shdl.mmu.edu.my/id/eprint/16492

Downloads

Edit (login required)