Potential Role of Convolutional Neural Network Based Algorithm in Patient Selection for DCIS Observation Trials Using...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Distinguishing pure Ductal Carcinoma In Situ (DCIS) from DCIS with invasion is crucial for appropriate patient management.
• Mammographic calcifications can indicate both non-invasive DCIS and occult invasive carcinoma.
• Accurate differentiation is essential to avoid under- or overtreatment.

Purpose of the Study:

• To evaluate the feasibility of using a convolutional neural network (CNN) for differentiating pure DCIS from DCIS with invasion based on mammographic images.
• To assess the diagnostic performance of the CNN model in this classification task.

Main Methods:

• A retrospective study utilized 246 mammographic images from 123 patients, categorized into pure DCIS and occult invasive groups.
• Calcifications were segmented and processed for a 15-hidden-layer CNN architecture with residual layers and dropout.
• The model underwent five-fold cross-validation using Keras with TensorFlow.

Main Results:

• The CNN model achieved an overall diagnostic accuracy of 74.6% (95% CI, ±5%).
• The area under the ROC curve was 0.71 (95% CI, ±0.04), with a specificity of 91.6% (95% CI, ±5%) and sensitivity of 49.4% (95% CI, ±6%).

Conclusions:

• Convolutional neural networks demonstrate feasibility in distinguishing pure DCIS from DCIS with invasion using mammographic data.
• The developed CNN model exhibits high specificity for this differentiation.