Answers are grounded in this paper and cite their sources.
Hepatic steatosis affects approximately 30.2% of the global population and, if undetected, can silently progress to non-alcoholic steatohepatitis, cirrhosis, and hepatocellular carcinoma. Manual CT interpretation for fatty liver assessment is time-consuming and prone to inter-observer variability, motivating research into automated screening tools. We propose RadDenseNet, a DenseNet121 model initialized with RadImageNet weights pretrained on 1.35 million medical images spanning CT, MRI, and ultrasound, and evaluate it as a research prototype. The first 397 of 427 backbone layers were frozen, and the remaining 30 layers were fine-tuned alongside a classification head comprising global average pooling, a dense layer of 256 units, batch normalization, and dropout of 0.5. The model was trained on the Kaggle Liver CT Image Dataset - 3,557 abdominal CT images across Healthy and Hepatic Steatosis classes - using a stratified 70/20/10 split with 5-fold cross-validation. RadDenseNet was compared against six baseline models spanning ImageNet and RadImageNet pretraining across ResNet50, InceptionV3, and DenseNet121 architectures. The decision threshold was optimized on the validation set, yielding τ = 0.48. On the held-out test set, RadDenseNet achieved 96.35% accuracy, AUC of 0.9966, sensitivity of 0.9538, specificity of 0.9752, precision of 0.9789, and F1-score of 0.9662, with 5-fold cross-validation confirming stable performance (mean accuracy 96.46% ± 0.68%, mean AUC 0.9962 ± 0.0014). RadImageNet pretraining consistently outperformed ImageNet pretraining across all architectures, with accuracy improvements of 2.81 to 5.90 percentage points. Grad-CAM analysis covering all four prediction categories revealed that false positives arose from extra-hepatic attention and false negatives from weak parenchymal activation - clinically meaningful failure modes. External validation on multi-center datasets, extension to multi-class steatosis grading, and liver ROI segmentation as preprocessing are recommended as future directions
No takes yet. Share an insight, caveat, or question.
Rahul et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: