A machine learning model based on MRI for the preoperative prediction of bladder cancer invasion depth

Abstract

Objectives: To construct and validate a prediction model based on full-sequence MRI for preoperatively evaluating the invasion depth of bladder cancer.

Methods: A total of 445 patients with bladder cancer were divided into a seven-to-three training set and test set for each group. The radiomic features of lesions were extracted automatically from the preoperative MRI images. Two feature selection methods were performed and compared, the key of which are the Least Absolute Shrinkage and Selection Operator (LASSO) and the Max Relevance Min Redundancy (mRMR). The classifier of the prediction model was selected from six advanced machine-learning techniques. The receiver operating characteristic (ROC) curves and the area under the curve (AUC) were applied to assess the efficiency of the models.

Results: The models with the best performance for pathological invasion prediction and muscular invasion prediction consisted of LASSO as the feature selection method and random forest as the classifier. In the training set, the AUC of the pathological invasion model and muscular invasion model were 0.808 and 0.828. Furthermore, with the mRMR as the feature selection method, the external invasion model based on random forest achieved excellent discrimination (AUC, 0.857).

Conclusions: The full-sequence models demonstrated excellent accuracy for preoperatively predicting the bladder cancer invasion status.

Clinical relevance statement: This study introduces a full-sequence MRI model for preoperative prediction of the depth of bladder cancer infiltration, which could help clinicians to recognise pathological features associated with tumour infiltration prior to invasive procedures.

Key points: • Full-sequence MRI prediction model performed better than Vesicle Imaging-Reporting and Data System (VI-RADS) for preoperatively evaluating the invasion status of bladder cancer. • Machine learning methods can extract information from T1-weighted image (T1WI) sequences and benefit bladder cancer invasion prediction.

Keywords: Bladder cancer; Machine learning; Magnetic resonance imaging; Medical imaging; Radiomics.