Performance Assessment of a Deep Learning-based Algorithm for Ovarian Cancer Histotyping in an Independent Data Set.

Area of Science:

• Oncology
• Pathology
• Artificial Intelligence

Background:

• Histotype classification in epithelial ovarian cancer is crucial for treatment but faces challenges due to inter-institutional slide variability.
• Domain adaptation techniques are needed to address variations in data from different sources.

Purpose of the Study:

• To evaluate the performance of the adversarial Fourier-based domain adaptation (AIDA) model in classifying five major ovarian cancer histotypes using an independent cohort.
• To assess the impact of retraining the AIDA model with additional slides on classification accuracy.

Main Methods:

• A retrospective study applied the AIDA deep learning model, trained on data from Vancouver General Hospital, to an independent cohort from Amsterdam University Medical Center.
• Histotype predictions were made using majority voting across 15 models, with retraining for misclassified cases using additional slides.
• Classification accuracy was assessed using single-slide and majority voting approaches.

Main Results:

• The AIDA model achieved an overall balanced accuracy of 79.7% across clear cell (CCC), endometrioid (EC), high-grade serous (HGSC), low-grade serous (LGSC), and mucinous (MC) ovarian cancer subtypes.
• Highest accuracies were observed for CCC (90.9%) and LGSC (89.8%), while EC (62.4%) showed the lowest.
• Retraining with additional slides improved balanced accuracy to 85.8% (single-slide) and 82.6% (majority voting).

Conclusions:

• The AIDA model demonstrates potential for accurate histotype classification in epithelial ovarian cancer, addressing domain shift challenges.
• Further model refinement is necessary to improve performance on difficult cases, potentially enhancing clinical diagnostic accuracy.