Ki-67 evaluation using deep-learning model-assisted digital image analysis in breast cancer

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Breast cancer diagnostics

Background:

• Accurate Ki-67 assessment is crucial for invasive breast carcinoma (IBC) prognosis.
• Manual Ki-67 scoring can be subjective and time-consuming.
• Digital image analysis offers potential for objective and efficient quantification.

Purpose of the Study:

• To evaluate the efficacy of artificial intelligence (AI)-assisted Ki-67 digital image analysis (DIA) in IBC.
• To quantitatively assess the performance of AI models in Ki-67 analysis.
• To determine the prognostic potential of AI-driven Ki-67 quantification.

Main Methods:

• Developed deep learning (DL) models, including a DL tissue classifier (DL-TC) and DL nuclear detection (DL-ND) model, using the HALO AI DenseNet V2 algorithm.
• Validated DL-TC performance using pixel-level ground truth annotations.
• Performed DL-TC and DL-ND-assisted DIA on 194 IBC cases, comparing results with manual Ki-67 counting and clinical outcomes.

Main Results:

• The DL-TC model demonstrated excellent performance in segmenting invasive carcinoma nests (average precision: 0.851; recall: 0.878; F1-score: 0.858).
• AI-assisted DIA showed strong positive correlations with manual Ki-67 index (ρ=0.961) and nuclei counts (ρ=0.928).
• High Ki-67 index (≥20%) was significantly associated with worse recurrence-free and breast cancer-specific survival (P=0.024, P=0.032).

Conclusions:

• The DL-TC and DL-ND models exhibited excellent performance.
• AI-assisted DIA provides a highly concordant assessment of Ki-67 compared to manual counting.
• This AI-driven approach demonstrates significant prognostic potential in invasive breast carcinoma.