Unified Calibrated Multi-Task NephroNet: Simultaneous Kidney Segmentation, Lesion Localization, and Pathological Classification in Abdominal CT Scans

• Ada John

  ladoke Akintola university of technology

  Author

Computed tomography (CT) remains the primary imaging modality for evaluating renal pathologies, yet existing deep learning approaches typically address kidney segmentation, lesion detection, and pathological classification as isolated tasks, resulting in redundant computation, fragmented clinical workflows, and miscalibrated confidence estimates that undermine clinical trust. This study proposes Unified Calibrated Multi-Task NephroNet (UCM-NephroNet), a multi-task deep learning framework that simultaneously performs kidney segmentation, lesion localization, and four-class pathological classification (normal, cyst, tumor, stone) on abdominal CT scans. The framework integrates a depthwise-separable encoder with task-specific decoders, employs homoscedastic uncertainty weighting for multi-task loss balancing, and applies post-hoc temperature scaling for probability calibration. Using a patient-disjoint, group-stratified evaluation protocol on 12,446 kidney CT images, UCM-NephroNet achieves 89.4% classification accuracy, Dice similarity coefficient of 0.94 for kidney segmentation, and lesion localization precision of 91.2%. Calibration analysis yields a Brier score of 0.018 and Expected Calibration Error of 0.023, substantially improving upon uncalibrated baselines. Feature importance analysis identifies peritumoral texture heterogeneity and contrast enhancement patterns as dominant predictors. The unified framework demonstrates that simultaneous multi-task learning with calibration-aware optimization produces clinically deployable outputs with trustworthy confidence estimates, offering a replicable template for integrated renal imaging diagnostics.

• pdf

Copyright (c) 2026 Ada John (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.