AI Summary & Key Takeaways
AI InsightAccuracy uncertainty was additionally quantified using 95% Wilson score confidence intervals.
📌 3 Key Takeaways
Abstract
Kidney CT imaging plays a central role in the urological diagnosis of renal cysts, calculi, solid tumors, and normal renal parenchyma. Although deep learning has shown promise for automated CT-based kidney pathology classification, existing approaches frequently rely on single-architecture models and lack systematic hyperparameter optimization, limiting both performance and reproducibility. This study aimed to develop and internally evaluate an optimized hybrid deep learning framework for multiclass kidney CT pathology classification in urological imaging, focusing on cyst, stone, tumor, and normal categories. A curated multiclass kidney CT dataset of 9,574 images was split into training (6,368), validation (1,997), and test (1,209) sets across four classes: Cyst, Normal, Stone, and Tumor. Four convolutional backbones-ResNet50V2, EfficientNetB3, DenseNet121, and Xception-were fine-tuned under a two-stage transfer learning protocol. Hyperparameters for each backbone were independently optimized using the Optuna framework with AdamW optimization and label smoothing. Class imbalance was addressed using weighted loss. Test-time augmentation was applied at inference. Two hybrid decision strategies were evaluated: validation-driven weighted soft voting and a logistic regression stacking learner. This was an internal computational validation study using a same-source dataset split; no external cohort was available. On the internal same-source held-out test set, ResNet50V2, Xception, the weighted ensemble, and the stacking model achieved perfect performance (100% accuracy, precision, recall, F1-score, and one-vs-rest ROC-AUC). DenseNet121 and EfficientNetB3 showed near-ceiling results. These outcomes may reflect strong class separability within the analyzed dataset and should be interpreted with caution in the absence of external validation. Ensemble strategies provided consistent performance gains over weaker individual backbones, particularly for the Stone class. The proposed optimized hybrid deep learning framework achieved strong internal performance for multiclass kidney CT pathology classification across urologically relevant categories. The findings suggest that architecture diversity and systematic hyperparameter optimization can substantially improve predictive robustness. External multi-center validation is required before clinical implementation in urological imaging workflows. The framework may serve as a candidate decision-support tool for radiologist-assisted interpretation of kidney CT findings, provided prospective validation is conducted. A revision-stage integrity audit found no byte-identical, pixel-identical, or repeated Roboflow source-identifier overlap across the predefined splits; however, two cross-split pairs showed very high structural similarity (SSIM >= 0.995), and patient-level independence could not be verified because patient identifiers were unavailable. Accuracy uncertainty was additionally quantified using 95% Wilson score confidence intervals.
Özet
Kidney CT imaging plays a central role in the urological diagnosis of renal cysts, calculi, solid tumors, and normal renal parenchyma. Although deep learning has shown promise for automated CT-based kidney pathology classification, existing approaches frequently rely on single-architecture models and lack systematic hyperparameter optimization, limiting both performance and reproducibility. This study aimed to develop and internally evaluate an optimized hybrid deep learning framework for multiclass kidney CT pathology classification in urological imaging, focusing on cyst, stone, tumor, and normal categories. A curated multiclass kidney CT dataset of 9,574 images was split into training (6,368), validation (1,997), and test (1,209) sets across four classes: Cyst, Normal, Stone, and Tumor. Four convolutional backbones-ResNet50V2, EfficientNetB3, DenseNet121, and Xception-were fine-tuned under a two-stage transfer learning protocol. Hyperparameters for each backbone were independently optimized using the Optuna framework with AdamW optimization and label smoothing. Class imbalance was addressed using weighted loss. Test-time augmentation was applied at inference. Two hybrid decision strategies were evaluated: validation-driven weighted soft voting and a logistic regression stacking learner. This was an internal computational validation study using a same-source dataset split; no external cohort was available. On the internal same-source held-out test set, ResNet50V2, Xception, the weighted ensemble, and the stacking model achieved perfect performance (100% accuracy, precision, recall, F1-score, and one-vs-rest ROC-AUC). DenseNet121 and EfficientNetB3 showed near-ceiling results. These outcomes may reflect strong class separability within the analyzed dataset and should be interpreted with caution in the absence of external validation. Ensemble strategies provided consistent performance gains over weaker individual backbones, particularly for the Stone class. The proposed optimized hybrid deep learning framework achieved strong internal performance for multiclass kidney CT pathology classification across urologically relevant categories. The findings suggest that architecture diversity and systematic hyperparameter optimization can substantially improve predictive robustness. External multi-center validation is required before clinical implementation in urological imaging workflows. The framework may serve as a candidate decision-support tool for radiologist-assisted interpretation of kidney CT findings, provided prospective validation is conducted. A revision-stage integrity audit found no byte-identical, pixel-identical, or repeated Roboflow source-identifier overlap across the predefined splits; however, two cross-split pairs showed very high structural similarity (SSIM >= 0.995), and patient-level independence could not be verified because patient identifiers were unavailable. Accuracy uncertainty was additionally quantified using 95% Wilson score confidence intervals.