Plexus Convolutional Neural Network for Histologic Imaging Analysis with Smaller Training Datasets and Parameters

Using a novel convolutional neural network architecture, PlexusNet can be used for histologic image analysis with smaller parameter and training sets than current state-of-the-art models. Reduction in parameter capacity prevents overfitting of the data, requires less data and is less computationally expensive. PlexusNet neural network architecture utilizes an optional normalization section with mandatory feature extraction and classification sections. In practice, a well-fitted PlexusNet model distinguished prostate cancer from healthy tissue on par with comparison models but was 23 times smaller and had better model calibration and clinical utility. Similarly, a model for detecting breast cancer metastases reduced required slide numbers by 43.8% and parameters by 200-fold.

Stage of Research