Machine learning based classification of aggressive and malignant renal tumors from multimodal data.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Renal tumors require accurate classification for effective treatment.
• Distinguishing between benign, malignant-indolent, and malignant-aggressive tumors is crucial for prognosis.
• Current classification methods can be enhanced by integrating imaging and clinical data.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) pipeline for classifying renal tumors.
• To assess the contribution of multiphase contrast-enhanced CT (CECT) images and clinical data in tumor classification.
• To differentiate between benign, malignant-indolent, and malignant-aggressive renal tumors.

Main Methods:

• A retrospective study included 448 patients with renal tumors.
• Multiphase CECT images underwent self-supervised feature extraction.
• Features were combined with clinical data and tumor size for classification using Random Forest (RF) and Multi-layer Perceptron (MLP) models.
• Nested five-fold cross-validation and AUC analysis were used for evaluation.

Main Results:

• The ML pipeline achieved an AUC of 0.90 for classifying indolent versus aggressive tumors.
• An AUC of 0.76 was achieved for classifying malignant versus benign tumors.
• Incorporating tumor size significantly improved classification accuracy, with RF excelling in indolent vs. aggressive and MLP in malignant vs. benign classification.

Conclusions:

• The developed ML pipeline accurately differentiates aggressive from indolent renal tumors, providing prognostic insights.
• Tumor size is a critical factor, enhancing the predictive value of CECT images and clinical data.
• ML techniques show significant potential for improving renal tumor risk stratification and personalized treatment.