StoneXAI-Net Classifies and Segments Kidney Stones Using Weakly Supervised CT Framework

Kidney stone disease, also known as urolithiasis or nephrolithiasis, represents one of the most common and recurrent urological disorders worldwide, imposing a significant clinical burden and substantial healthcare costs across healthcare systems. Unenhanced non-contrast computed tomography serves as the primary diagnostic imaging modality for evaluating suspected urolithiasis, providing high sensitivity and specificity across varying calculus sizes, anatomical locations, and mineral compositions. However, building automated artificial intelligence systems to reliably identify, delineate, and triage renal calculi from abdominal computed tomography scans has historically faced steep computational and clinical hurdles. Standard deep learning pipelines for medical image segmentation typically depend on massive volumes of manual, pixel-level annotations drafted by expert radiologists, which remain prohibitively time-consuming and expensive to acquire at scale. Furthermore, abdominal imaging presents intricate radiological challenges: adjacent skeletal structures such as the vertebral column and ribs, as well as contrast-enhanced parenchyma and bowel, display hyperdense attenuation values that easily confuse naive machine learning algorithms into generating severe false-positive localizations. Compounding this challenge, explainability methods in computer vision are frequently applied in a circular fashion, relying on the same saliency signals to supervise weak localization and subsequently to validate model faithfulness.

To overcome these systemic bottlenecks, computer science researchers P. Karthick and Chiranji Lal Chowdhary from the School of Computer Science Engineering and Information Systems at Vellore Institute of Technology developed StoneXAI-Net. Introduced in a study published in Frontiers in Artificial Intelligence, the framework offers an explainable, weakly supervised multi-task system engineered to classify, localize, and segment kidney stones directly from axial computed tomography slices without requiring manually contoured pixel masks. Instead of relying on expensive expert segmentations, StoneXAI-Net couples physics-grounded training data augmentation with anatomically constrained consensus pseudo-labeling, an attention-prior-guided Attention U-Net architecture, multi-stone detection capabilities, and a decoupled interpretability audit. The primary clinical motivation behind the architecture is not merely achieving elevated benchmark metrics, but satisfying stringent trustworthiness and safety requirements at the clinical bedside, ensuring that diagnostic software correctly distinguishes calculi from non-stone hyperdensities, behaves robustly across scans containing multiple stones, and provides verifiable explanations that withstand rigorous quantitative audits rather than relying on qualitative visual heatmaps.

The developmental framework was evaluated using a benchmark dataset of 3,364 original axial abdominal computed tomography images collected from three medical facilities, comprising Mercy, Faruq, and Rania hospitals in the Sulaimani and Rania regions. This public imaging cohort contains 1,577 slices presenting confirmed kidney stones and 1,787 normal slices displaying no calculi. All images were prepared as grayscale slices normalized to a unit intensity range and standardized to a resolution of 224 by 224 pixels. To ensure rigorous experimental methodology and completely eliminate data leakage across evaluation folds, the authors partitioned the cohort using five-fold cross-validation, strictly isolating a held-out test and validation split of 672 original images that never received synthetic augmentation. Within the remaining training split of 2,692 slices, the classes were initially balanced to 1,430 images per category before undergoing a specialized curriculum of image transformations designed to expand the effective training volume without compromising validation integrity.

Unlike conventional computer vision methods that apply generic geometric transformations such as basic rotations, flips, and scaling, the initial module of StoneXAI-Net incorporates a specialized computed tomography physics-guided augmentation suite. Conventional transformations fail to capture the underlying radiological physics of X-ray computed tomography, including quantum photon statistics, beam hardening artifacts, and variable Hounsfield display window settings, which can impair cross-scanner robustness. To bridge this physical gap, the researchers modeled low-dose quantum photon fluctuations using Poisson noise distributions and simulated beam hardening alongside calcium blooming artifacts by introducing targeted Gaussian blooming centered on hyperdense calculus proxies while removing oriented dark streaks. Furthermore, the pipeline introduces Hounsfield window and level jitter within calibrated gain and level intervals, accompanied by a stone-aware CutMix operator. This CutMix technique samples auxiliary patches from donor slices and composites them into background regions while strictly avoiding any overlap with hyperdense stone voxels. Applied online with an effective multiplier factor of four, this physics-driven suite quadrupled the balanced training partition to 11,440 images, creating an even split of 5,720 stone images and 5,720 normal images that exposed the learning backbones to realistic scanner artifacts.

The classification backbone of the architecture utilizes an ImageNet-pretrained ResNet-50 network, fine-tuned specifically across its third and fourth residual stages and final fully connected classification layers while retaining earlier frozen representations to curb overfitting. Mini-batches were processed with stone-aware CutMix training, optimizing parameters through the Adam optimizer using a learning rate of 0.0005 under five-fold cross-validation with an early stopping mechanism monitoring held-out validation slices. Across all five folds, the classifier demonstrated remarkable diagnostic stability, securing an aggregate accuracy of 97.4 percent, a precision of 98.4 percent, a sensitivity recall of 96.0 percent, an F1-score of 97.1 percent, and an area under the receiver operating characteristic curve of 99.5 percent. Statistical reliability confirmed through one thousand bootstrap resamples revealed narrow 95 percent confidence intervals spanning 0.967 to 0.979 for accuracy and 0.992 to 0.996 for the area under the curve, accompanied by standard deviations below one percent across folds. The model exhibited a consistently conservative diagnostic profile where precision outpaced recall, yielding 63 false negatives compared to only 26 false positives across individual image evaluations.

To generate reliable spatial supervision without human annotators, the best-performing classification models drive a multi-engine consensus pseudo-labeling pipeline. Weakly supervised localization often suffers when relying on single class activation maps, which tend to concentrate narrowly on small discriminative centers and easily trigger on non-stone skeletal features. The StoneXAI-Net architecture simultaneously computes three distinct activation algorithms from the final residual stage: higher-order gradient-based Grad-CAM++, forward-pass confidence-driven Score-CAM, and perturbation-based Ablation-CAM. The pipeline establishes consensus by averaging the normalized maps while computing pixel-wise standard deviations to quantify spatial disagreement. Pixels displaying high disagreement among the three algorithms are suppressed by an uncertainty agreement gate calibrated at an empirical threshold of 0.35. The remaining consensus field is smoothed using conditional random fields or Gaussian filtering, providing an initial probability map that highlights candidate stone regions based on collective multi-algorithmic agreement.

Recognizing that raw saliency heatmaps cannot reliably differentiate dense bone from actual renal calculi, the authors introduced an explicit four-stage anatomical constraint module that operates as a strict spatial prior. The anatomical filter first extracts the outer body boundary to eliminate external air and the examination table, then identifies the prominent hyperdense spinal structure positioned posteriorly and centrally in the supine axial scan. By establishing spatial references relative to the midline vertebral body, the algorithm reconstructs an anatomically bounded retroperitoneal renal search envelope scaled to body radius while excluding skeletal landmarks, spinous processes, laminae, and ribs. Candidate structures inside this renal search zone are evaluated using an explicit scoring metric incorporating activation attention, raw Hounsfield intensity, local contrast, and morphological compactness, while rejecting highly elongated structures with eccentricity above 0.97 as non-calculus vascular or costal artifacts. Systematic threshold sensitivity analysis across a dedicated development cohort revealed that an operating threshold of 0.55 produced the optimal trade-off, achieving an 81.5 percent stone yield alongside a reduced normal-sample false-positive rate of 33.5 percent. When deployed across the dataset, 405 generated pseudo-masks, representing 81.8 percent of candidate outputs, passed the high-confidence gate to train downstream segmenters, successfully rejecting 90 uncertain or anatomically implausible slices.

The segmentation phase employs an Attention U-Net that receives a two-channel input combining the original grayscale computed tomography slice with the anatomically constrained consensus prior map. This architecture incorporates additive attention gates at each decoder level, where coarser feature representations dynamically gate skip-connection signals to prevent spatial gradients from wandering into irrelevant tissue. To accelerate structural learning on tiny stone targets, the decoder integrates two auxiliary deep-supervision heads operating at intermediate resolutions. Training proceeds under an 85 to 15 split of the high-confidence pseudo-labeled data, guided by a multi-component hybrid loss combining soft-Dice loss for regional overlap balance, focal loss to counter severe foreground-background disparity, Sobel edge boundary loss for precise margin delineation, and an attention-prior consistency penalty. In proxy validation against consensus pseudo-masks, the complete Attention U-Net configuration achieved an average proxy Dice similarity coefficient of 0.834, an Intersection over Union of 0.777, and a 95th-percentile Hausdorff distance of 2.94 pixels. Notably, incorporating anatomical post-processing reduced the Hausdorff boundary error by 6.26 pixels compared to an unconstrained baseline, demonstrating that anatomical priors effectively anchor the predicted contours to true renal boundaries.

Beyond single-stone segmentation, StoneXAI-Net includes a multi-stone localization and detection engine designed to handle complex bilateral or multi-focal clinical presentations. Extracted components thresholded at a probability of 0.5 are scored through a multiplicative cascade combining mean segmentation probability, a hyperdense voxel fraction scaled to prevent premature discarding of low-density calculi, and the primary classification confidence score. In multi-stone validation, the admission rule successfully retained confirmed secondary and tertiary calculi across 59 out of 495 slices, representing 11.9 percent of cases, without collapsing multiple distinct calculi into single detections or misidentifying adjacent skeletal structures. When mapped through free-response receiver operating characteristic analysis across sixteen operational confidence thresholds, the detection module achieved a peak per-stone sensitivity of 0.875 while maintaining exceptionally low false-positive rates between 0.00 and 0.03 false alarms per scan, highlighting the practical utility of cascade gating for clinical triage.

A critical methodological advance in StoneXAI-Net is its decoupled explainability audit, specifically engineered to eliminate the widespread flaw of circular validation in machine learning research. In conventional weakly supervised studies, investigators frequently use class activation maps to supervise pseudo-mask generation and subsequently present those same activation heatmaps as evidence that the network is interpretable and faithful. StoneXAI-Net systematically breaks this confirmation loop by assessing post-hoc explainability exclusively through Integrated Gradients, an axiomatic path-attribution technique entirely decoupled from the Grad-CAM++, Score-CAM, and Ablation-CAM algorithms used during training. The audit quantified explanation fidelity using deletion curves, insertion curves, energy concentration ratios, and pointing-game accuracy within a 15-pixel radius. On quantitative testing, Integrated Gradients achieved an insertion score of 0.835, a pointing-game accuracy of 0.81, and an energy concentration of 0.103, verifying that the underlying network focus aligns objectively with true stone morphology rather than reflecting an artifact of the supervisory saliency signal.

While StoneXAI-Net establishes an innovative benchmark for weakly supervised medical imaging, the authors transparently emphasize several vital methodological and clinical limitations. Because the underlying public dataset contains slice-level labels without comprehensive patient identifiers, the experimental evaluation could not implement strictly patient-independent cross-validation, meaning current findings should be interpreted as an exploratory proof-of-concept. Most importantly, because independent expert radiologist pixel annotations were unavailable, all reported segmentation and localization metrics represent proxy assessments of consistency against consensus pseudo-masks rather than clinically confirmed ground-truth accuracy. Furthermore, the framework was developed on a single retrospective public dataset without external multi-center validation across diverse scanner manufacturers and imaging protocols. Looking toward future clinical translation, the investigators aim to expand the StoneXAI-Net architecture from two-dimensional axial slices to full three-dimensional volumetric computed tomography analysis, enabling volumetric stone burden quantification, precise spatial sizing, and external multi-center validation backed by expert radiological annotations.

Subject of Research: Weakly supervised deep learning framework for automated kidney stone analysis on computed tomography

Article Title: StoneXAI-Net: an explainable weakly supervised framework for kidney stone classification, localization, and segmentation from CT images

Article References: Karthick, P., & Chowdhary, C. L. (2026). StoneXAI-Net: an explainable weakly supervised framework for kidney stone classification, localization, and segmentation from CT images. Frontiers in Artificial Intelligence, 9, Article 1966841. https://doi.org/10.3389/frai.2026.1966841

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1966841

Keywords: StoneXAI-Net, kidney stone disease, urolithiasis, computed tomography, weakly supervised learning, Attention U-Net, consensus pseudo-labeling, CT-physics augmentation, explainable artificial intelligence, Integrated Gradients, bone exclusion, multi-stone localization, free-response ROC, ResNet-50, medical image segmentation, axial CT imaging