Project Link: https://github.com/justinvyu/siemens-2017
PyTorch, NumPy, Matplotlib
Esophageal cancer (EC) is a highly lethal malignancy worldwide with a 5- year survival rate below 20%. Accurate diagnostic tools predicting clinical outcomes and disease progression are desperately needed. We developed PathoNet, a novel deep-learning-based diagnostic software that automates immunohistochemistry scoring. PathoNet was uniquely designed with four steps: (1) formatting images into trainable tiles, (2) passing tiles through FilterNet, a convolutional neural network, and (3) ExpressNet, another convolutional neural network; and (4) aggregating tile scores to a final score. Instead of using packaged pre-trained models, we created our FilterNet and ExpressNet using the open-source PyTorch library, modeling after AlexNet architecture. PathoNet is currently optimized to score E-Cadherin (PathoNet-E-Cad), a biomarker that may predict EC progression and overall survival. Trained with 3072 tiles, PathoNet scores showed 85.62% tile-level concordance and 91.67% image-level concordance with pathologists, outperforming published automated immunohistochemistry scoring systems. We demonstrated the clinical potential of PathoNet-E-Cad by testing on 473 patient samples. The PathoNet-E-Cad score is associated with esophageal disease progression. Low PathoNet-E-Cad score is significantly correlated with better overall survival (p=0.043) and predicts optimal treatment outcomes of EC (p=0.027). More biomarkers are being integrated into PathoNet to further facilitate EC prognosis.