Domain-adaptive semi-supervised learning for efficient rare pathological lesion detection with minimal annotation.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Pathology
• Computational Pathology

Background:

• AI for rare pathological lesion detection faces challenges due to scarce expert annotations and domain shifts across institutions.
• Performance degradation is significant, with up to 70.3% reduction in detection precision for rare kidney lesions.

Purpose of the Study:

• To develop and evaluate a robust AI methodology for rare pathological lesion detection across multi-institutional kidney biopsies.
• To address domain shifts caused by different scanner types and institutional variations.

Main Methods:

• Utilized multi-institutional kidney biopsy data from 22 hospitals with three scanner types (NDPI, VSI, SVS).
• Integrated semi-supervised learning with residual CycleGAN-based domain adaptation.
• Evaluated context-dependent optimal strategies for different scenarios.

Main Results:

• Reduced mean Fréchet inception distance between institutions from 55.9 to 20.2, preserving diagnostic morphology.
• Semi-supervised learning improved rare lesion detection by 15.2-17.7% in same-hospital scenarios.
• Combined GAN-Semi-Supervised approach improved detection by up to 63.4% for crescents in cross-scanner scenarios (NDPI vs. VSI).

Conclusions:

• The proposed methodology enables robust AI performance for rare pathological lesion detection across diverse healthcare settings.
• This approach minimizes the need for extensive expert annotation while enhancing model generalizability.
• Context-specific strategies optimize AI performance depending on data variability (intra- vs. inter-institutional).