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

摘要: Rationale and Objectives We investigated the feasibility of utilizing convolutional neural network (CNN) for predicting patients with pure Ductal Carcinoma In Situ (DCIS) versus DCIS invasion using mammographic images. Materials Methods An IRB-approved retrospective study was performed. 246 unique images from 123 were used our CNN algorithm. total, 164 in 82 diagnosed by stereotactic-guided biopsy calcifications without any upgrade at time surgical excision (pure group). A total 41 yielding occult invasive carcinoma as final upgraded diagnosis on surgery (occult Two standard magnification views (CC ML/LM) analysis. Calcifications segmented an open source software platform 3D Slicer resized to fit a 128 × 128 pixel bounding box. 15 hidden layer topology implement network. The architecture contained five residual layers dropout 0.25 after each convolution. Five-fold cross validation performed training set (80%) (20%). Code implemented Keras TensorFlow Linux workstation NVIDIA GTX 1070 Pascal GPU. Results Our algorithm achieved overall diagnostic accuracy 74.6% (95% CI, ±5) area under ROC curve 0.71 ±0.04), specificity 91.6% ±5%) sensitivity 49.4% ±6%). Conclusion It's feasible apply distinguish high