Hybrid I-ResNeT-ViT and Cost-Sensitive InceptionV3 Models for Tumour Severity and Malignancy Classification Using Medical Imaging.

backgroundEarly tumour severity and malignancy classification are critical for timely clinical intervention, improving patient outcomes through accurate risk assessment. However, conventional diagnostic methods often struggle to accurately differentiate tumour severity and malignancy at early stages, frequently leading to misdiagnosis or delayed intervention.

objectiveThe aim is to develop a robust deep learning framework for accurate, automated, and interpretable classification of tumour severity and malignancy across multiple medical imaging modalities.

methodsHigh-resolution mammography, MRI, and CT images were collected from publicly available repositories, ensuring representation of diverse tumour types, stages, and malignancy levels. Data pre-processing involved resizing, noise reduction, and Histogram Equalization with Region-Based Segmentation (HE-RBS) to enhance image contrast and isolate regions of interest. Feature extraction utilized a hybrid iResNet with ViT Feature Fusion (iRViT-HFF) to capture both local and global tumour characteristics. Tumour severity and malignancy were classified using an Explainable Cost-Sensitive InceptionV3 (CS-InceptionV3) model to minimize critical misclassifications and provide clinically interpretable outputs.

resultsThe proposed framework achieved 97.6% accuracy, 96.9% sensitivity, 98.3% specificity, and a 97.2% F1-score, significantly outperforming conventional machine learning and other deep learning methods. The model reliably classified early-stage tumours across all imaging modalities and provided interpretable heatmaps to support clinical decision-making.

conclusionThe hybrid deep learning framework accurately and effectively classifies tumour severity and malignancy at early stages, providing a reliable and interpretable tool to support clinical oncology workflows.