Research Overview (Non-Technical)


On November 15, 2019, a post appeared on Twitter answering the following question: Who do we need in healthcare today? (figure on the left hand side) Healthcare is in need of professionals (rigorously) trained in the theoretical and computational aspects of machine learning and systems engineering, radiology imaging informatics, and who hold patient care much much higher than "singular technological aspirations". This ideology is rigorously practiced in the Algorithms in Theory and Practice (ATP) Lab at LUMS. As the name of my research lab suggests, we study fundamental properties of algorithms, its computational virtues, limitations, and challenges, and its efficacy in completing a clinical task in question. Our focus is always on solving a clinical problem or completing a clinical task successfully i.e., we value effectiveness over elegance. In fact, we strive to develop a compact and explainable computational algorithm (elegance) which robustly and accurately solves a clinical problem (effective). At ATP Lab, my students and I are working on several projects in computational clinical imaging expounded below.

The RANO-AI Program

RANO-AI is an acronym for Radiomics (and Radiogenomics) in Neuro-Oncology using AI. The ultimate goal of the RANO-AI program is to develop a computational imaging pipeline for brain tumor characterization using radiological imaging data only. We focus on brain gliomas which account for 30% of all brain and central nervous system tumors and 80% of all malignant brain tumors. The RANO-AI program focuses on the following problems:
  • Automatic Segmentation: Designing a deep segmentation network (including architecture, components, configuration, and training routine) for fully automatic, robust, accurate, and generalizable segmentation of tumor subregions on 3D MRI scans.
  • Segmentation with limited annotations: Designing a learning framework for training deep segmentation networks with limited manually annotated data while optimally exploiting abundant unlabeled (non-annotated) datasets.
  • Feature Stability: Identifying features which are not robust to non-physiological variations in the MRI scans (including scanner effects, reconstruction protocols, post-processing), feature extraction software, and deep segmentation algorithms.
  • Prospective evaluation: Thoroughly evaluating the proposed computational imaging pipeline with prospective studies..
  • Fairness evaluation: Detailed examination of the computational imaging pipeline to identify potential sources of bias and heuristic shortcuts which are misconstrued by the learned model as a physiological signature.
  • Foundation Models: We are also exploring the integration of Foundation Models/LLMs/Vision-Language models in the RANO-AI program. This is a new and vibrant domain of research in the AI community and we are also engrossed in discovering innovative ways of integration – where multi-modal information can be exploited for guided training of deep models.

Semi-supervised Learning in Cardiac Image Analysis

ATP Lab is also pursuing research endeavors in cardiology. We are developing computational algorithms, to be integrated in a pipeline, for radiomics in cardiac imaging. The goal is robust and accurate prediction of cardiac pathologies from 3D MRI (or CT) scans only and quantification of clinical measures of interest, which define left-ventricle and right-ventricle function. We are currently working on a novel semi-supervised segmentation algorithm for multi-phase 3D cardiac image segmentation. This will be followed by an extensive radiomics study with 3D cardiac MRI (or CT) scans.

Topological Image Analysis

ATP Lab has embarked on a new research frontier which explores application of topological data analysis tools in medical image analysis (Topological Image Analysis). The challenge lies in discovering novel ways of integrating topological tools within a deep learning framework (including neural network architecture and loss function). Topological Data Analysis is an advanced subject with a strong mathematical component (theory) and a rich computational component (algorithms). We are also leading a Journal Club on Topological Data Analysis. Participants include graduate students, doctoral students, research assistants/associates, and faculty.