Pathology Image Analysis Laboratory

The Mahmood Lab aims to utilize machine learning, data fusion, and medical image analysis to develop streamlined workflows for cancer diagnosis, prognosis, and biomarker discovery. We are interested in developing automated and objective mechanisms for reducing interobserver and intraobserver variability in cancer diagnosis using artificial intelligence as an assistive tool for pathologists. The lab also focuses on the development of new algorithms and methods to identify clinically relevant morphologic phenotypes and biomarkers associated with response to specific therapeutic agents. We develop multimodal fusion algorithms for combining information from multiple imaging modalities, familial and patient histories and multi-omics data to make more precise diagnostic, prognostic and therapeutic determinations.

Deep Learning for Medical Image Analysis

Our research focuses on deflating the hype around artificial intelligence (AI) for medical imaging and developing unique solutions for addressing pressing issues with the applicability of deep learning methods for medical image analysis. These include:

  • Synthetic Data Generation: Using synthetic data to overcome the limited availability of annotated and labeled medical images.
  • Domain Adaptation: Deep networks trained on synthetic data often fail to adapt to real data, my work focuses on using adversarial training and domain randomization for adapting networks trained on synthetic data to real data.
  • Context-Aware Deep Learning: Classical deep learning-based approaches rely on a pixel-wise loss function, my research focuses on using joint deep learning-graphical model setups and adversarial training to make deep networks more context-aware.
  • Multimodal Feature Fusion: Physicians often make a diagnosis based on a number of different imaging modalities, prior knowledge, patient and familial histories, etc. My work focuses on fusing information from multiple modalities as well as modeling clinical intuition into deep networks for improved performance and more accurate diagnosis.
  • Localizing Features used for Classification: Deep learning often enables diagnosis with a high degree of accuracy, localizing the features used to make classification decisions can lead to the discovery of new biomarkers.

Multimodal Data Fusion for Improved Diagnosis, Prognosis, and Response to Treatment

Our work focuses on data-driven determination and subtyping of cancer and consequently predicting survival and treatment outcomes. This often requires integration and fusion of information obtained from multiple data-types. Molecular characterization often forms the basis of diagnosis and targeted therapy. But molecular information alone is rarely used to make such therapeutic determinations and is often complemented by histopathology and radiology data. We envision developing data-driven multimodal fusion methods for improved diagnosis, prognosis, and response to specific therapeutic agents.


Data-driven Biomarker and Morphological Feature Discovery