Development of a deep learning-based model to evaluate changes during radiotherapy using cervical cancer digital...

Area of Science:

• Digital pathology
• Machine learning in oncology
• Cervical cancer research

Background:

• Accurate classification of cervical cancer biopsies is crucial for treatment planning and prognosis.
• Radiotherapy response assessment often relies on subjective interpretation of histopathological features.
• Developing objective, quantitative methods for analyzing tumor response is an unmet clinical need.

Purpose of the Study:

• To develop and validate a deep learning model for classifying cervical cancer biopsies before and during radiotherapy.
• To visualize model predictions on whole slide images (WSIs) for better interpretability.
• To explore the clinical significance and prognostic value of the model's output, termed radiotherapy status probability (RSP).

Main Methods:

• Utilized DenseNet121 feature extractor and support vector machine classifier for a deep learning model.
• Trained and tested the model on approximately 12,400 and 6,000 hematoxylin-eosin stained biopsy tiles, respectively.
• Assessed model performance per-tile and per-WSI, visualizing RSP as a color map on WSIs and performing survival analysis.

Main Results:

• The model achieved an area under the receiver operating characteristic curve (AUC) of 0.76 per-tile and 0.95 per-WSI on the test set.
• Visualization highlighted viable tumor components and stroma.
• While RSP during treatment lacked prognostic impact, low RSP at diagnosis was associated with prolonged overall survival (P=0.045).

Conclusions:

• A deep learning model was successfully developed to classify cervical cancer biopsies before and during radiotherapy, with results visualized on WSIs.
• The model's ability to analyze tumor morphologic features shows potential for predicting patient prognosis.
• Cases with lower RSP before treatment demonstrated a better prognosis, indicating the potential utility of these AI-derived features.