ISSP 2030 Spring AI Forum Week 7

12 Mar 2021 - Arun Balajiee

Date: 03/12/2021

Speaker: Yanwu Xu, DBMI

In his talk, Xu talked about the implementation of a system for brain tumor segmentation from medically processed images using the approaches from Transfer Learning, Domain Adaptation and Domain Shift. Specifically, in terms of weakly labeled data. His system is split in two stages. In the first stage, training a network on a source domain and target domain.In the second stage, this network is then adapdated to work on the target domain using a disciminator to further imrpove the performance of the tumor detection. From the results the model performs better than state-of-art model for unsupervised learning approaches on weakly labeled data. The baseline model was a network trained only on the source domain. More details of his implemtnation are available in his publication.

The keytakeways from the talk are that domain adaptation and shift with transfer learning, use of a pre-trained network and fine-tuning on smaller datasets, can greatly improve the performance of the current state-of-art models, which only use unsupervised approaches or weakly-labeled datasets alone to train systems for brain tumor segementation and classification from medical images.

Speaker: Giacomo Nebbia, DBMI

In this talk, Nebbia talked about the implementation of a multi-task learning approach to breast cancer detection from medical images. Specifically, his research questions were to find out if additional information about the patient, not just the labeling of the images themselves, during training could improve model performance and if a transfer learning approach that have failed previously in the classification of medical images for cancerous or benign tumors that incorporate clinical data perform better than models that only use pre-trained CNNs such as ImageNet, GoogleNet and ResNet. The datasets for Nebbia’s model uses the quantitative and qualitivate labeling from target domain incorportated in two ways – Concurrent training of knowledge labels and curriculum learning of knowledge lables. For the multi-task learning the Nebbia’s framework offers two different outputs for two tasks – diangons of cancer and BI-RADS diagnosis score. In concurrent rating appraoch, the model is trained for these tasks in parallel. In curriculum knowledge -guided training, the clinical information is used to train the network. Both approaches outperform the pre-trained CNNs used on the same dataset.

Some interesting questions discussed were, the possibilities of comparing this with the Inter-Rater Reliability Scores (IRR) of two different radiologists’ diagnosis on the same images and if there was an upper-bound on the performance of the model as is seen in most cases of computational biology. As a scope of future, this model will be transfer learned on different images of cancerous tissues in another organs, with datasets that images of people from different demographics.