Deep Learning-Based Multi-Scale Attention Fusion Framework for Automated Kidney Cancer Localization and Classification: Kidney Cancer
Year: 2026
Type: article
Abstract: Background: CT imaging is one of the most commonly used methods for diagnosing renal malignancy. Manual evaluation is cumbersome and highly subjective as it depends on the skill set of an expert nephrologist. Methods: This paper proposes an automated deep learning approach based on a two-stage model in lesion localization and early-stage kidney cancer classification. First, the kidney regions are segmented using an improved U-Net architecture with multi-scale feature fusion, enabling accurate identification of affected kidney regions. Next, the segmented lesion maps are fed into a classifier network (ResNet/DenseNet) to classify early-stage kidney cancer. Results: Segmentation-based lesion information fusion improves stage-wise discrimination and limits false predictions during the early stages of kidney cancer detection. The proposed method clearly demonstrates improved predictive accuracy and operational efficiency compared to conventional end-to-end classification models. The integration of segmentation and classification provides more reliable identification of cancerous regions and supports improved stage-wise prediction. Conclusion: The approach provides clinical interpretability by marking the affected portion of the kidney and assigning a stage-wise cancer label. The proposed framework can assist clinicians in early-stage kidney cancer assessment by providing both lesion localization and stage-wise classification. It also offers a promising basis for developing efficient and interpretable computer-aided diagnostic systems for renal malignancy.
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7218656719
