Fig. 4.

Illustration of the partial transfer learning scheme. “Conv”, “Max pooling”, “ReLU” and “Full” denote the convolutional layer, max pooling layer, rectified linear unit layer and fully connected layer, respectively. The network was trained on the ImageNet data containing millions of labeled natural images with thousands of categories (left). The pre-trained parameters are partially transferred to the target domain of biological images. In particular, we truncated the pre-trained CNN model at layer 21, and attached one max pooling and two fully connected layers to obtain the new CNN model. Then we used multi-task learning approach to fine-tune the modified CNN model using labeled ISH images (right).