Abstract
In Europe, endometrial carcinoma was found to be the fourth most common tumor in the female population in 2022. The depth of myometrial invasion is a well-established and crucial prognostic risk factor in endometrial cancer patients and is important for choosing the most appropriate management for the patient. However, while the preoperative assessment of tumor invasion carried out using radiological imaging is very important, it is a subjective examination and its accuracy is based on the experience of the operator. In this scenario we proposed a radiomic-based model to predict myometrial invasion for ultrasound images. We collected clinical data and qualitative ultrasound indicators of 77 consecutive patients affected by endometrial carcinoma. After a pre-processing phase of ultrasound images, a pre-trained Inception-V3 convolutional neural network (CNN) was used as features extractor. Then, a binary classification model and a multiclass model were trained, after a double step of feature selection; the first selection stage performed feature filtering based on a nonparametric test, the second stage selected features by evaluating not only the relationship with the outcome of interest, but also the relationship with other predictive features. For the multiclass prediction model, a cascade approach has been developed. The two proposed models were validated in 100 ten-fold cross-validation rounds. In addition, to assess the effect of the potential added value of using this tool in clinical practice, we compared the operator’s performance with and without the developed automated support. The binary and multiclass model reached optimal classification performances with a mean AUC value equals to 90.76 (88.63–92.89 IC95%). When the operator was assisted by the radiomic-based decision-making system increased by 10% points in terms of precision. The multiclass model showed the per-classes recall were 93.33%, 71.88% and 90.00%, for focal infiltration, infiltration less than 50%, and infiltration greater than 50% class, respectively. The performances of the operator when assisted by the radiomic-based decision-making system were statistically superior both in terms of overall accuracy and per-class recall of intermediate class (rising to 82.82% and 71.88% compared to 71.88% and 56.25%, respectively). The proposed models have the potential to standardize examinations that rely on subjective evaluations, such as ultrasound. They can represent a valid support tool for the clinicians for an accurate estimate of the deep myometrial infiltration already in the diagnosis phase with an easily accessible, low-cost examination that causes no discomfort for the patient such as ultrasound.