Diagnostic Accuracy StudyEditorially reviewed

Preoperative CT-based deep learning radiomics model to predict lymph node metastasis and patient prognosis in bladder cancer: a two-center study.

Rui Sun, Meng Zhang, Lei Yang, Shifeng Yang, Na Li, Yonghua Huang, Hongzheng Song, Bo Wang, Chencui Huang, Feng Hou, Hexiang Wang

Insights into Imaging · January 25, 2024 · Vol 15 · Issue 1 · p. 21

PMID: 38270647PMCID: PMC10811316DOI: 10.1186/s13244-023-01569-5

Open original article

This is an evidence-grounded AI reading aid that has passed human editorial review. Items markedRadWeave readingare our interpretation rather than statements by the authors. It is not a substitute for reading the original publication, and RadWeave does not reproduce the source full text here.

Take-home message

A CT-based combined model integrating radiomics and clinical features shows promise for predicting lymph node metastasis in bladder cancer, outperforming the clinical model alone on external testing, though validation remains limited.

RadWeave reading

Conventional CT lymph node assessment in bladder cancer has well-documented sensitivity limitations, and this radiomics-augmented approach is positioned as a potential complement—not yet a replacement—for standard staging.

RadWeave reading

What the radiologist should know

Current CT and MRI size-based lymph node assessment in bladder cancer has low sensitivity, meaning nodal metastases are frequently missed at staging; this study attempts to address that gap with a radiomics-based model.
Evidence grounded · 2 source references
The combined model is a nomogram incorporating a LightGBM radiomics signature—using hand-crafted and deep learning features—alongside stalk presence and CT-reported lymph node status as clinical predictors.
Evidence grounded · 3 source references
The CT-based combined model predicted lymph node metastasis status in bladder cancer patients on external testing, with the radiomics signature alone achieving higher discrimination but the combined nomogram achieving higher accuracy and better decision-curve utility than the radiomics signature alone.
Evidence grounded · 3 source references
The clinical model, built from stalk presence and CT-reported lymph node status identified by multivariate logistic regression, showed notably lower discrimination on external testing than on the training set, illustrating the limited generalizability of low-dimensional visual features alone.
Evidence grounded · 2 source references
Combining hand-crafted radiomics with deep learning features and applying SMOTE oversampling to address class imbalance may be a methodological approach worth considering in future radiomics studies where lymph-node-positive cases are a minority.
RadWeave reading
Evidence grounded · 2 source references

Reporting implications

This paper does not support a change to reporting.

No established routine reporting change is supported. Feature awareness: Model is investigational; external test set is small, survival stratification did not hold in external testing, and multiparametric MRI/prospective validation are lacking. No established routine reporting change is supported.

Practice impact

RadWeave editorial assessment, not a statement by the authors.

Potentially useful

The combined model shows promising discrimination for nodal staging in a setting where CT sensitivity is known to be poor. However, the small external cohort, retrospective design, and absent prospective validation limit immediate clinical translation.

Evidence grounded · 3 source references

Caveats before applying this

The combined model achieved significant progression-free survival risk stratification in the total cohort and training set but not in the external test set, which the authors attributed to selection bias: a much higher proportion of lymph-node-positive patients in the external set had prolonged survival compared with the training set.
Evidence grounded · 3 source references
Manual ROI segmentation was used throughout; this is time-consuming and introduces inter-observer variability that could reduce reproducibility if the approach were applied in routine practice.
RadWeave reading
Evidence grounded · 1 source reference
Data came from only two centers and required harmonization; performance in centers with different CT protocols or patient populations is unknown and may be lower than reported.
RadWeave reading
Evidence grounded · 1 source reference
The study is retrospective and limited to CT; the authors themselves note that multiparametric MRI and larger multicenter prospective validation are needed before broader clinical use.
Evidence grounded · 1 source reference

Numbers worth remembering

Combined model — AUC: 0.834 (95% CI: 0.659–1.000)
Clinical model — AUC: 0.764 (95% CI: 0.697–0.831)
Clinical model — AUC: 0.624 (95% CI: 0.402–0.846)
Combined model — Accuracy: 0.870
Radiomics signature — Accuracy: 0.852

RadWeave bottom line

This investigational CT radiomics nomogram shows promising discrimination for lymph node metastasis in bladder cancer but requires prospective, multicenter validation before any change in routine reporting practice can be considered.

RadWeave reading

Evidence grounded · 4 source references

Study in 20 seconds

Study type: Diagnostic Accuracy Study
Population: Bladder cancer patients who underwent three-phase CT and surgical resection with extended pelvic lymph node dissection, from two centers
Modality: Three-phase CT (primary lesion segmentation; lymph node imaging deliberately excluded)
Technique: LightGBM radiomics signature combining hand-crafted and deep learning features with SMOTE oversampling, integrated into a nomogram with clinical predictors
Comparator / reference: Clinical model (stalk presence and CT-reported lymph node status) and radiomics signature alone
Primary endpoint: Prediction of lymph node metastasis status (AUC and accuracy on external test set)