Robustness of Nomogram Versus Deep Learning Models for Hepatocarcinoma Grading: Insights From a Bilateral...

This study aimed to compare the performance of nomograms and deep learning (DL) models in predicting the pre-operative pathological grade of hepatocellular carcinoma (HCC) using clinical data and ultrasound images, with an emphasis on model robustness and generalizability across institutions.

We retrospectively analyzed 2404 pre-operative ultrasound images from 601 patients with pathologically confirmed HCC across two medical centers. Two distinct modeling strategies were employed: two nomograms integrating ultrasound imaging features (B-mode and contrast-enhanced ultrasound [CEUS]) with clinical variables, and a series of DL models based on B-mode, CEUS at arterial phases, portal phases and delayed phases, and a combined B + CEUS modality. Independent predictors of pathological grade were identified using univariate, Least Absolute Shrinkage and Selection Operator and logistic regression and incorporated into the nomograms. To rigorously assess generalizability, bilateral training and a cross-center validation strategy were implemented: data from Center 1 were split 7:3 for training and internal validation, while all data from Center 2 served as an external test set. This process was reversed (i.e., training on Center 2 and testing on Center 1) to complete the bilateral cross-center validation. Model performance was evaluated using the area under the curve (AUC), calibration curves and decision curve analysis.

The nomograms demonstrated strong and consistent performance in predicting HCC pathological grade, with AUCs of 0.771 (internal validation) and 0.762 (external validation) when trained on Center 1 data, and AUCs of 0.814 (internal) and 0.818 (external) when trained on Center 2 data. In contrast, DL models showed high internal performance (AUCs: 0.632-0.809 for Center 1; 0.692-0.767 for Center 2) but experienced significant degradation on external validation (AUCs: 0.599-0.646 and 0.620-0.657, respectively).

The nomogram exhibited superior robustness and generalizability compared to DL models in heterogeneous clinical populations. These findings highlight the practical advantage of interpretable statistical models in settings with variable data distributions and reinforce the value of feature-based approaches when data diversity is limited.