An interpretable model based on weakly supervised learning for uterine smooth muscle tumor diagnosis: A multi-center...

Area of Science:

• Gynecologic Pathology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Uterine smooth muscle tumors (USMTs) present diagnostic challenges due to overlapping morphology between benign leiomyomas (LM), malignant leiomyosarcomas (LMS), and smooth muscle tumors of uncertain malignant potential (STUMP).
• Accurate differentiation is critical for appropriate patient management, but histological evaluation can be complex and time-consuming.

Purpose of the Study:

• To develop and validate a weakly supervised artificial intelligence (AI) model for classifying USMTs as benign or malignant using only whole-slide image (WSI) level labels.
• To assess the model's diagnostic performance and interpretability, and evaluate its utility in AI-human collaboration for pathology tasks.

Main Methods:

• A multi-center dataset of 94 LMS and 634 benign USMT cases (total 1797 WSIs) was used for training and internal testing.
• An independent external test set of 27 LMS and 90 benign USMT cases (total 302 WSIs) was utilized for validation.
• A CAMEL2-based weakly supervised learning model was developed, and its diagnostic performance was evaluated using Area Under the Curve (AUC) and accuracy metrics. Interpretability was assessed via heatmaps, and an AI-human collaboration study was conducted.

Main Results:

• The AI model achieved excellent diagnostic performance, with AUCs of 0.9976 (internal) and 0.9889 (external), and accuracies exceeding 0.97.
• Interpretability heatmaps highlighted key pathological features of LMS in malignant cases and suspicious regions in STUMP cases, aligning with pathologist annotations.
• In AI-human collaboration, model assistance led to improved diagnostic accuracy and reduced diagnostic time.

Conclusions:

• This study presents the first weakly supervised AI model for USMT diagnosis, demonstrating high accuracy and interpretability with minimal annotation.
• The AI model shows significant potential as a decision-support tool for pathologists, enhancing diagnostic efficiency and accuracy in gynecologic pathology.
• Weakly supervised learning offers a promising approach for developing AI tools in pathology, overcoming limitations of extensive data annotation requirements.