A deep learning-based automatic staging method for early endometrial cancer on MRI images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early diagnosis of endometrial cancer (EC) significantly improves patient survival rates.
• Deep learning (DL) offers advanced computer-aided diagnosis for medical image analysis, potentially reducing radiologist misdiagnosis.
• Accurate staging of EC is crucial for effective early treatment planning.

Purpose of the Study:

• To develop and validate an automated deep learning (DL) model for early-stage endometrial cancer (EC) staging using magnetic resonance imaging (MRI).
• To assess the efficacy of DL-based segmentation and tumor-to-uterus ratio (TUR) analysis for differentiating between EC stages IA and IB.
• To provide an effective, automated staging tool for radiologists to aid in early EC diagnosis.

Main Methods:

• Retrospective analysis of 117 early EC patients' MRI scans (axial T2WI, axial DWI, sagittal T2WI).
• U-net based DL model for semantic segmentation of uterine and tumor regions.
• Calculation of tumor-to-uterus ratio (TUR) and receiver operating characteristic (ROC) curve analysis for staging.

Main Results:

• The DL segmentation model achieved high accuracy (Dice coefficients: 0.910-0.972).
• Classification models demonstrated strong performance with Area Under the Curve (AUC) values of 0.86 (axial T2WI), 0.85 (axial DWI), and 0.94 (sagittal T2WI).
• The automated DL approach effectively differentiated between EC stages IA and IB.

Conclusions:

• An automated DL-based segmentation model combined with TUR analysis on MRI is an effective method for early endometrial cancer staging.
• This DL approach can assist radiologists in improving the accuracy and efficiency of EC diagnosis.
• The findings support the clinical utility of DL in enhancing early cancer detection and patient management.