International Journal of Science, Engineering and TechnologyAuthors: Aminu Jafar, Abubakar Muhammad Miyim, Zahraddeen Sufyan, Nuraddeen Jaafar Ibrahim
Abstract: Radiomics has emerged as a promising paradigm for transforming standard medical images into high-dimensional quantitative features that can support early cancer detection and treatment planning. However, clinical adoption of radiomics-based artificial intelligence (AI) has been hindered by computational inefficiency, long inference times, and limited generalizability across diverse imaging settings. This research addresses these challenges by developing a hybrid CNN–Vision Transformer (ViT) pipeline combined with Principal Component Analysis (PCA) and Light Gradient Boosting Machine (LightGBM) for the diagnostic imaging of breast, lung, and colorectal cancers. A multi-institutional dataset of 6,360 cases was curated from The Cancer Imaging Archive (TCIA), MICCAI repositories, and two Nigerian hospitals (ABUTH and UCH). Radiomic features (n = 1,834) were extracted in compliance with the Image Biomarker Standardisation Initiative (IBSI). CNN and ViT models were used for complementary feature extraction (local and global contexts), followed by feature fusion and dimensionality reduction with PCA. The reduced features (150 principal components, retaining 99.2% variance) were classified using LightGBM. The system was evaluated using Accuracy, Sensitivity, Specificity, F1-score, and AUC-ROC, with statistical significance assessed via DeLong, McNemar, and bootstrap confidence intervals. The proposed model achieved 94.2% overall accuracy, with per-cancer accuracies of 94.7% (breast), 92.3% (lung), and 93.8% (colorectal), and AUC values ≥ 0.923. Sensitivity and specificity averaged 0.93 and 0.92 respectively. Crucially, the pipeline achieved an average inference latency of 1.7 ± 0.3 seconds per image, representing a >99% reduction in computational time compared to baseline CNN (4.1 minutes) and ViT (3.2 minutes) models. Ablation studies confirmed the incremental value of each component: CNN-only accuracy (87.2%), ViT-only (89.4%), hybrid without PCA/LightGBM (91.3%), versus full hybrid (94.2%). Explainability analysis using SHAP identified texture heterogeneity, shape irregularity, and intensity statistics as the most influential feature families, improving model transparency and clinical trust. The findings demonstrate that the proposed hybrid radiomics-AI model simultaneously delivers state-of-the-art diagnostic accuracy and near real-time inference speed, directly addressing key barriers to clinical adoption. Its validated performance across international and Nigerian cohorts highlights strong generalizability, and its interpretability makes it suitable for deployment in real-world oncology imaging workflows. This work establishes a new benchmark for high-speed, explainable AI-driven radiomics in cancer diagnostics.
DOI: http://doi.org/10.5281/zenodo.17498894
