I want to work on a binary classification problem using a medical dataset consisting of 35K images. I have a few questions on the same.
1.) The architectures like VGG, Inception, etc which are typically used on datasets like COCO, ImageNet, etc are modeling input images that are "rich" and "complex" in features. VGG has around 26 million parameters and InceptionV4 around 50 million, does it make sense to use such deep models for medical images?
2.) I know of multiple architectures that are employed for multi-class classification problems but are there any neural network architectures that are specifically designed for binary classification instead of converting just the last layer?
3.) Any suitable references on how to deal with class imbalance problem which typically arises in binary classification for medical datasets? I mean medical datasets typically pose this as sort of a combination of binary classification and anomaly detection. The one way I have found is to use SVM instead of the last fully connected layer but any better methods?
