We propose a framework for learning feature representations for variable-sized regions of interest (ROIs) in breasthistopathology images from the convolutional network properties at patch-level. The proposed method involvesne-tuning a pre-trained convolutional neural network (CNN) by using small xed-sized patches sampled from theROIs. The CNN is then used to extract a convolutional feature vector for each patch. The softmax probabilitiesof a patch, also obtained from the CNN, are used as weights that are separately applied to the feature vectorof the patch. The final feature representation of a patch is the concatenation of the class-probability weightedconvolutional feature vectors. Finally, the feature representation of the ROI is computed by average poolingof the feature representations of its associated patches. The feature representation of the ROI contains localinformation from the feature representations of its patches while encoding cues from the class distribution of thepatch classication outputs. The experiments show the discriminative power of this representation in a 4-classROI-level classication task on breast histopathology slides where our method achieved an accuracy of 66:8% ona data set containing 437 ROIs with dierent sizes.
展开▼
