Image-intensity normalization approaches on T1-weighted MRI images for lipomatous soft tissue tumors

<h2>Abstract</h2><h3>Purpose</h3> To propose new image-intensity normalization approaches to overcome image-intensity variations on magnetic resonance imaging (MRI) and compare the effects of these approaches on radiomic feature reproducibility. <h3>Material and method</h3> This retrospective study included 70 patients diagnosed with intramuscular lipoma and atypical lipomatous tumors/well-differentiated liposarcoma, confirmed via histopathology, who underwent preoperative MRI examinations between 2010 and 2020. Two observers segmented the region of interest for all tumors on T1-weighted images (T1WIs) and extracted radiomic features from the normalized images using four different methods (<mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mi>Z</mml:mi></mml:mrow></mml:math>, <mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mrow><mml:mi>N</mml:mi><mml:mi>I</mml:mi><mml:mi>V</mml:mi></mml:mrow></mml:mrow></mml:math>, <mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mi>M</mml:mi></mml:mrow></mml:math>, and <mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mrow><mml:mi>F</mml:mi><mml:mi>M</mml:mi></mml:mrow></mml:mrow></mml:math>). We evaluated the performance of the methods in terms of the stable feature rate and the area under the receiver operating characteristic curve (AUC) of the machine-learning model on the reproducible features. We compared the stable feature rates and the AUCs between the models using the chi-squared test and the Delong's test, respectively. <h3>Results</h3> The extraction process yielded a total of 107 radiomic features, including 14 shape, 18 first-order, and 75 texture features. The proposed method <mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mrow><mml:mi>F</mml:mi><mml:mi>M</mml:mi></mml:mrow></mml:mrow></mml:math> exhibited a higher stable feature rate than the other methods (<mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mrow><mml:mi>F</mml:mi><mml:mi>M</mml:mi></mml:mrow></mml:mrow></mml:math>, 86.0 % vs. <mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mi>Z</mml:mi></mml:mrow></mml:math>, 26.2 %; <mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mrow><mml:mi>N</mml:mi><mml:mi>I</mml:mi><mml:mi>V</mml:mi></mml:mrow></mml:mrow></mml:math>, 74.8 %; and <mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mi>M</mml:mi></mml:mrow></mml:math>, 49.5 %; all <i>P</i> < 0.05). The AUC of the model for the <mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mrow><mml:mi>F</mml:mi><mml:mi>M</mml:mi></mml:mrow></mml:mrow></mml:math> method was 0.95 (0.86–1), higher than that for the other methods. <h3>Conclusions</h3> The proposed method <mml:math><mml:mrow><mml:mi>I</mml:mi><mml:mrow><mml:mi>F</mml:mi><mml:mi>M</mml:mi></mml:mrow></mml:mrow></mml:math> can reduce T1WI intensity variations. This approach should be evaluated in future multicenter studies to develop the machine-learning model in predicting intramuscular lipomas and atypical lipomatous tumors/well-differentiated liposarcomas.