Mehdi Moradi  Mehdi Moradi photo         

contact information

Scientist/Manager, AI for Healthcare
IBM Almaden Research Center, San Jose, CA, USA


A significant portion of my work and the work of my team focuses on deep learning for medical image analysis, in scenarios where data is limited, unbalanced, or without labels. Here are a few recent contributions. 

Learning in data limited scenarios:

In recent years, deep neural networks have improved the benchmarks of learning in areas such as vision and speech recognition. This improvement comes with a big price tag. Deep neural networks are very large supervised learning models and need huge quantities of labelled data at the time of training. In medical image analysis, labeling data is expensive. We work towards reducing the burden of data labeling with two approaches:

AI for AI: producing labels through learning

Most of clinical data exists in unstructured formats without labels. This poses a huge challenge to data hungry supervised deep learning approaches that have revolutionized AI. This work introduces a new technology that uses images and reports often available in PACS systems and produces local annotations on images that highlight the region of the image depicting diseases and findings. This allows building large scale locally annotated image databases for deep learning and enable the use of new AI methods in a vast array of clinical applications.

Relevant publication: Mehdi Moradi, Ali Madani, Yaniv Gur, Yufan Guo, Tanveer Syeda-Mahmood, “Bimodal network architectures for automatic generation of image annotation from text”, Proceedings of MICCAI 2018, pp. 449-456.

Learning with fewer labeled samples: semi-supervised learning with a generative adversarial architecture

In many scenarios, unlabeled data is abundant, but labeled data is scarce. In this work we propose a new architecture for classification of chest X-ray images that can use unlabeled data. The idea is that unlabeled data can contribute to learning the representation of data. The discriminatory features of normal and diseased CXRs are minor variations that the classifier can learn from a small labeled set, as long as the CXR image domain is learnt. The implementation of this ideas takes advantage of a modified generative adversarial net where the discriminator has output nodes for fake, real-normal, and real disease labels. The loss function is modified to handle fake, unlabeled and labeled CXR images differently.  When training is complete, the discriminator classifier is used as the disease classifier. We show an order of magnitude reduction in the required size of labelled dataset with this architecture, as compared to a traditional convolutional neural network.


Relevant publication: Ali Madani, M. Moradi, Tanveer Syeda-Mahmood, “Semi-supervised learning with generative adversarial networks for chest X-ray classification with ability of data domain adaptation”, In IEEE International Symposium on Biomedical Imaging (IEEE ISBI), pp. 1038-1042, 2018.


Learning from normal: Curriculum learning

Given the huge variety of possible clinical conditions in an imaging modality, such as chest computed tomography (CT), it is extremely challenging to build a sufficiently large dataset with samples of abnormalities. As a result, most learning-based medical image analysis solutions focus on a narrow range of diseases. Apart from lack of generality, limited positive samples also create unbalanced datasets. If such datasets are directly used to train classifiers, low true positive rates can be expected. Given these limits, we propose a new strategy for building medical image analysis pipelines that target disease detection. In simple terms, we try to extract useful knowledge from the negative samples and make use of such knowledge to improve classification on limited positive samples. Inspired by the idea of curriculum learning, we propose a strategy for building medical image classifiers using features from segmentation networks. By using a segmentation network pre-trained on similar data as the classification task, the machine can first learn the simpler shape and structural concepts before tackling the actual classification problem which usually involves more complicated concepts. Using our proposed framework on a 3D three-class brain tumor type classification problem, we achieved 82% accuracy on 191 testing samples with 91 training samples. When applying to a 2D nine-class cardiac semantic level classification problem, we achieved 86% accuracy on 263 testing samples with 108 training samples.

Relevant publication: Ken C. L. Wong, Tanveer Syeda-Mahmood, Mehdi Moradi, “Building medical image classifiers with very unbalanced and limited data using segmentation networks”, Medial Image Analysis, vol. 49, pp. 105-116, 2018.

Relevant publication: Ken C. L. Wong, Alexandros Karargyris, Tanveer Syeda-Mahmood, and M. Moradi, “Building Disease Detection Algorithms with Very Small Numbers of Positive Samples.” MICCAI 2017, LNCS 10435, pp. 471-479, 2017.


For more information on topics of this page, contact Mehdi Moradi, PhD (