Semi-supervised learning in prostate MRI tumor detection approaches fully supervised performance on external...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Developing AI tools for prostate cancer detection on MRI is hindered by the need for extensive expert annotations.
• Semi-supervised learning offers a potential solution to reduce this annotation burden.

Purpose of the Study:

• To evaluate the diagnostic performance of semi-supervised learning models for aggressive prostate cancer detection on MRI.
• To compare semi-supervised models against fully supervised models trained with additional expert annotations.

Main Methods:

• Utilized 1500 MRI scans from the PI-CAI challenge dataset.
• Compared a novel teacher-student semi-supervised approach (mtU-Net) against supervised and semi-supervised nnU-Net models.
• External validation performed on PROMIS and Prostate158 datasets, evaluating performance using AUC and AP.

Main Results:

• Fully supervised nnU-Net achieved the highest performance on internal and external validation datasets.
• The proposed semi-supervised mtU-Net demonstrated comparable external validation performance to the fully supervised model.
• Semi-supervised methods significantly outperformed the supervised baseline on external datasets.

Conclusions:

• Fully supervised learning yields the best performance for prostate MRI tumor detection.
• Semi-supervised learning approaches demonstrate performance close to fully supervised models in external validation.
• Semi-supervised learning is a valuable strategy when expert annotations are limited, facilitating scalable AI diagnostic tools.