MRI-based Ovarian Lesion Classification via a Foundation Segmentation Model and Multimodal Analysis: A Multicenter...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Artificial intelligence (AI) shows potential in enhancing diagnostic accuracy for ovarian lesions on MRI.
• However, the generalizability of AI models across diverse datasets remains uncertain.
• Developing efficient and robust AI pipelines is crucial for clinical application.

Purpose of the Study:

• To develop an efficient and generalizable AI pipeline for ovarian lesion characterization using MRI.
• To integrate automated segmentation with deep learning classification for improved diagnostic performance.
• To compare the AI pipeline's performance against expert radiologists.

Main Methods:

• Retrospective analysis of multiparametric MRI datasets from three institutions.
• Automated lesion segmentation using Meta's Segment Anything Model (SAM).
• Training and external validation of a DenseNet-121 deep learning (DL) model incorporating imaging and clinical data.

Main Results:

• SAM-assisted segmentation reduced processing time by 4 minutes per lesion with high Dice coefficients (0.86-0.88).
• The DL model achieved an area under the receiver operating characteristic curve (AUC) of 0.85 internally and 0.79 externally.
• AI classification performance was comparable to that of radiologists (AUC: 0.84-0.93, P > .05).

Conclusions:

• An accurate and efficient AI pipeline integrating SAM and DL for ovarian lesion classification on MRI was developed.
• The pipeline demonstrated reduced segmentation time and performance comparable to radiologists.
• This approach offers a promising tool for differentiating benign from malignant ovarian lesions.