Multi-Level Cancer Classification with FlexiFocusNet Using a Novel Nigerian Computed Tomography Imaging Dataset and Standard Histopathology Data.
Authors
Affiliations (8)
Affiliations (8)
- School of International Education, Oxford Brookes College, Chengdu University of Technology, Chenghua District, Chengdu, Sichuan, 610059, China. [email protected].
- College of Nuclear Technology and Automation Engineering and Sichuan Engineering Technology Research Center for Industrial Internet Intelligent Monitoring and Application, Chengdu University of Technology, Chenghua District, Chengdu, Sichuan, 610059, China. [email protected].
- Network and Data Security Key Laboratory of Sichuan Province, University of Electronic Science and Technology of China, Chengdu, Sichuan, China. [email protected].
- Network and Data Security Key Laboratory of Sichuan Province, University of Electronic Science and Technology of China, Chengdu, Sichuan, China.
- Radiology Department, Federal Medical Centre, Umuahia, Abia State, Nigeria.
- Life Elixir Computers, Owerri, Imo State, Nigeria.
- Faculty of Pharmaceutical Sciences, University of Nigeria Nsukka, Nsukka, Enugu State, Nigeria.
- Department of Food Science and Technology, Madonna University, Akpugo, Enugu State, Nigeria.
Abstract
Reliable and generalisable artificial intelligence-based cancer image analysis remains challenging because of data scarcity, population bias, and the need to capture both local morphological details and long-range contextual dependencies. Many existing deep learning approaches rely on single-scale feature extraction or isolated attention mechanisms, potentially limiting their representation of multi-level spatial and channel information across imaging tasks. To address these challenges, we introduce FlexiFocusNet, an engineering-driven deep learning framework for multi-level cancer classification across heterogeneous medical imaging modalities. The architecture integrates hierarchical feature extraction, multi-scale attention fusion, transformer-style self-attention, and feature refinement to combine fine-grained local features with global contextual information. FlexiFocusNet was evaluated on a newly curated Nigerian computed tomography (CT) dataset for binary and three-class cancer classification and on a standard five-class histopathology dataset to examine its performance across imaging modalities. For binary CT classification, the model achieved 99.30 ± 0.49% accuracy, 99.62 ± 0.43% balanced accuracy, and an AUC of 1.000 under five-fold cross-validation, with 98.04% accuracy, 98.09% balanced accuracy, and an AUC of 0.999 on the independent test set. For three-class CT classification, it achieved 98.36 ± 1.05% validation accuracy and 98.01 ± 1.12% balanced accuracy, with 98.83% test accuracy, 98.52% balanced accuracy, and an AUC of 0.996. On five-class histopathology data, FlexiFocusNet achieved 99.92 ± 0.05% validation accuracy and 99.89 ± 0.06% balanced accuracy, followed by 100% test accuracy and an AUC of 1.000. The consistently high performance across the evaluated tasks indicates effective feature learning across different classification settings and imaging modalities. The study also contributes an underrepresented Nigerian CT dataset, subject to applicable institutional data-sharing requirements, providing an additional resource for research on cancer image analysis and geographically diverse medical artificial intelligence.