Speaker 1: Yanwu Xu
Title: Box-Adapt: Domain-Adaptive Medical Image Segmentation using Bounding Box Supervision
Speaker 2: Giacomo Nebbia
Title: Multi-task learning to incorporate clinical knowledge into deep learning for breast cancer diagnosis
Summary and thoughts:
Both the talks presented at this week’s ISP forum were related to transer learning and applications of deep learning in medicine. While Yanwu Xu presented his domain adaptation framework with the application of liver segmentation in mind, Giacomo Nebbia’s deep learning approach focused on breast cancer diagnosis. Transfer learning has become a popular approach in deep learning applications and plays an important role in several fields - from biomedicine to self-driving cars. The idea involves transfer of knowledge from one task or domain to another. This is typically done by learning parameters or features in one task and transferring the knowledge to a different domain. The goal is generally to achieve a low prediction error on the target task. The motivation behind application of transfer learning can vary depending on the problem task. Reasons may include unavailability of sufficient data for target domain and incorporation of additional knowledge for auxiliary tasks.
Both speakers presented their work with transfer learning applications, albeit with different approaches. Yanwu Xu presented his framework called Box-Adapt, a weakly-supervised domain adaptation method which performs joint training on the source and target domains in two stages and then self-trains with labels of the target domain. The self-adaptation on the target domain serves to overcome limitations of traditional domain adaptation which needs sufficient labeled data on the target dataset. Giacomo Nebbia’s talk presented his work on domain knowledge incorporation in deep learning through incorporation of clinical knowledge from radiologists in addition to the available medical images in the model for breast cancer diagnosis and biopsy outcome prediction. Evaluation and comparison with baseline models on two relevant datasets showed that incorporating clinical knowledge improves performance of CNN models and a combination of knowledge and ImageNet pretraining is superior to domain knowledge alone. Both talks effectively presented the utility and application of transfer learning with different goals and schemes. While proliferation of data in some fields allows traditional deep learning algorithms to perform well, there are many applications where availability of data is still low and transfer learning approaches can be used to develop models for accurate prediction.