Can Only the Shape Feature in Radiomics Help Machine Learning Show That Bladder Cancer Has Invaded Muscles?

OBJECTIVES

The presence of muscle invasion is an important factor in establishing a treatment strategy for bladder cancer (BCa). The aim of this study is to reveal the diagnostic performance of radiomic shape features in predicting muscle-invasive BCa.

METHODS

In this study, 60 patients with histologically proven BCa who underwent a preoperative MRI were retrospectively recruited. The whole tumor volume was segmented on apparent diffusion coefficient (ADC) maps and T2W images. Afterward, the shape features of the volume of interest were extracted using PyRadiomics. Machine learning classification was performed using statistically different shape features in MATLAB (The MathWorks, Inc., Natick, Massachusetts, United States).

RESULTS

The findings revealed that 27 bladder cancer patients had muscle invasion, while 33 had superficial bladder cancer (53 men and seven women; mean age: 62±14). Surface area, volume, and relevant features were significantly greater in the invasive group than in the non-invasive group based on the ADC maps (P<0.05). Superficial bladder cancer had a more spherical form compared to invasive bladder cancer (P=0.05) with both imaging modalities. Flatness and elongation did not differ significantly between groups with either modality (P>0.05). Logistic regression had the highest accuracy of 83.3% (sensitivity 82.8%, specificity 84%) in assessing invasion based on the shape features of ADC maps, while K-nearest neighbors had the highest accuracy of 78.2% (sensitivity 79.1%, specificity 69.4%) in assessing invasion based on T2W images.

CONCLUSIONS

Shape features can be helpful in predicting muscle invasion in bladder cancer using machine learning methods.