Importances of hormone receptors
More than 2 million women worldwide were diagnosed with breast cancer in 2018, resulting in 0.6 million deaths.
A large majority of invasive breast cancers are hormone receptor positive: tumor cells grow in the presence of estrogen (ER) and/or progesterone (PR). The presence of these receptors determines the diagnosis and especially the treatment given to patients.
In clinical practice, pathologists diagnose breast cancer from hematoxylin and eosin staining and then estimate estrogen receptors from Immunohistochemistry staining.
ReceptorNet
ReceptorNet is a model, developed by Salesforce Research and Dr. David Agus of the Lawrence J. Ellison Institute for Transformative Medicine at the University of South Carolina, able to predict the status of hormone receptors.
ReceptorNet shows that a deep learning algorithm can accurately predict the presence of these receptors directly from hematoxylin and eosin staining.
ReceptorNet is able to predict hormone receptor status from an inexpensive tissue image. This contrasts with the current standard of care (Immunohistochemistry), which requires both a more expensive and less widely available type of tissue image and a trained pathologist to review these images.
ReceptorNet is based on ResNet-50
The ReceptorNet architecture consists of three interconnected neural networks: a feature extractor module, an attention module and a decision module.
The feature extractor is a ResNet-50 without the softmax layer, followed by two fully connected layers with a dropout of 0.5 converting the 1000-dimensional features obtained from the ResNet-50 into a 512-dimensional feature vector.
Adding additional layers to a sufficiently deep neural network first leads to accuracy saturation and then to accuracy degradation. The ResNet architecture solves this problem. ResNet consists of residual blocks.
Model training and performance
The investigators used data containing a total of 3399 hematoxylin and eosin images of patients provided by the Australian Breast Cancer Tissue Bank (ABCTB) and The Cancer Genome Atlas (TCGA). The dataset shows wide variation in preparation and scanning quality which is better for the generalization of the model.
ReceptorNet is trained using input patches sampled directly from WSI (whole slide images), without pixel-level annotations. It automatically learns to pay attention to the regions of the WSI that are important for the estimation of hormone receptors.
The output of the model is a binary classification: ER+or ER-.
To develop this model, Salesforce Research researchers used the PyTorch language and the computing power of Nvidia P-100 Pascal graphics processing units (GPUs). ReceptorNet achieves an area under the curve (AUC) of 0.92 for sensitivity and specificity.
ReceptorNet Predictions
ReceptorNet automatically learns that low-grade tumors (a), invasive lobular carcinoma (b), reactive stroma and carcinoma in situ (d) are predictive of ER+; and high-grade tumors (e) and necrotic debris (f) are predictive of ER-.
ReceptorNet also learns to ignore adipose tissue, connective tissue with few or no tumor cells, and macrophages loaded with debris and fat (g).
Know more about ResNet
