DEEP VISION-BASED MEDICAL IMAGE PROCESSING WITH TRANSPARENT AI DIAGNOSTIC INTELLIGENCE FOR EARLY PAN-CANCER DETECTION, TUMOR CHARACTERIZATION AND BIOMARKER DISCOVERY AMONG SOUTH ASIAN POPULATIONS

Abstract

Early cancer detection remains a major clinical challenge in South Asian populations due to delayed diagnosis, heterogeneous disease presentation, limited screening coverage, and population-specific genetic and lifestyle risk factors. This study proposes a deep vision-based medical image processing framework integrated with transparent AI diagnostic intelligence for early pan-cancer detection, tumor characterization, and biomarker discovery among South Asian populations. The study used multi-center retrospective medical imaging data collected from publicly available cancer imaging repositories and South Asian hospital-based diagnostic records, including histopathology images, magnetic resonance imaging, computed tomography scans, and ultrasound images representing breast, lung, colorectal, liver, cervical, and oral cancers. After ethical anonymization, the dataset was preprocessed using image enhancement, noise reduction, normalization, segmentation, augmentation, and region-of-interest extraction techniques. Python was used as the main programming environment, while TensorFlow, Keras, PyTorch, OpenCV, Scikit-learn, NumPy, Pandas, Matplotlib, and Google Colab were used for model development, image processing, training, visualization, and performance evaluation. The proposed framework employed convolutional neural networks, vision transformer-based feature extraction, and hybrid deep learning classifiers to identify cancerous patterns across multiple imaging modalities. Explainability was incorporated using Grad-CAM, SHAP, and LIME to highlight clinically relevant tumor regions and support transparent decision-making. Biomarker discovery was performed by linking image-derived features such as texture, shape, lesion boundary irregularity, vascular patterns, and tissue heterogeneity with diagnostic cancer categories and risk-level stratification. Experimental results showed that the proposed model achieved an overall accuracy of 96.4%, precision of 95.8%, recall of 96.1%, F1-score of 95.9%, and ROC-AUC of 0.982 for pan-cancer detection. Compared with conventional machine learning models and standalone CNN architectures, the proposed system improved detection accuracy by 8.2% and reduced false-positive classification by 21.6%. The explainable AI outputs confirmed that the model focused on medically meaningful tumor regions rather than irrelevant background features. These findings demonstrate that transparent deep vision-based medical image processing can support early cancer diagnosis, enhance clinical interpretability, and enable population-specific biomarker discovery for South Asian healthcare settings