Javid
Dadashkarimi

I am a Postdoctoral Research Fellow at MGH/Harvard, working on developing deep learning models for MRI image segmentation with a focus on overcoming data scarcity and leveraging synthetic data. I completed my Ph.D. at Yale University under the guidance of Professors Dustin Scheinost and Amin Karbasi, with research on applying optimal transport to connectomics data. For more details, check out the latest version of my CV.

click this art to regenerate it
I whipped up this little piece of art that changes every time you click on it.

With a passion for AI-driven solutions in medical imaging, I focus on developing cutting-edge tools that elevate diagnostic precision and advance patient care. My research has been featured in some of the leading journals and conferences, including Nature, Molecular Psychiatry, MICCAI, and COLING. Church Doors In addition to medical imaging, I bring expertise in natural language processing, particularly in machine translation and information retrieval, where I also hold a patent. You can explore my personal projects on GitHub. I am currently available for freelance opportunities in prototyping, consulting, and technical advisory roles. For collaboration inquiries, feel free to reach out – let’s connect!

Also here is my updated Linkedin profile, academic X page, Google Scholar.

Acronyms Explained

DL Deep Learning
NLP Natural Language Processing
MRI Magnetic Resonance Imaging
ECG Electrocardiography
AF Atrial Fibrillation
CLIR Cross Lingual Information Retrieval
YINS Yale Institute of Network Sciences

Selected Projects

The tools and solutions I’ve built are versatile and designed for deployment across various data modalities, including time-series data, stock market analysis, EEG signals for neural data, cardiovascular ECG data for heart disease prediction, and complex imaging data such as fMRI and MRI.

Specifically, I focus on brain segmentation, body organ segmentation, skull stripping, brain and organ masking, and face detection. In areas like prenatal neuroimaging, where training data is limited, I employ generative AI to produce synthetic images and precise masks, enabling robust model development in these challenging domains. Church Doors

In one of our projects, we developed a sliding-window approach specifically designed to improve automated fetal brain extraction from full-uterus MRI images. Church Doors This method addresses the common issue of false-positive tissue predictions by navigating towards the fetal brain and selectively extracting brain tissue. Using a breadth-fine search (BFS) followed by a deep-focused sliding window (DFS), our approach refines the extraction process, significantly reducing false positives. Trained on a small set of synthetic fetal brain images, our model achieves state-of-the-art accuracy, especially in challenging second-trimester cases and for EPI scans.

Some of my projects are:

Title Institution Date Topic Info
Cascade U-Nets MGH/Harvard 2024/10/27 DL for Segmentation

This project implements a hybrid 3D, multi-scale model for fetal brain extraction from MRI scans using a cascade of U-Net architectures. Using a Breadth-Fine Search (BFS) and Deep-Focused Sliding Window (DFS) approach, our framework achieves precise segmentation results in full-uterus stack-of-slices scans, even with limited training data and annotations.

Our method employs a cascade of four models—A, B, C, and D—trained to focus on 3D patches of decreasing sizes from a synthesized set of training images augmented with random geometric shapes. At inference, we pool sliding-window predictions across multiple sizes to detect and refine the region of interest (ROI) that most likely contains the fetal brain. Subsequently, we derive an accurate brain mask through voxel-wise majority voting.

Pipeline Overview

In the BFS stage, Models A and D generate probability maps \( P_A \) and \( P_D \) across the full input volume \( I \), producing a coarse brain mask from the largest connected components across maps. We then perform a bounding box fit on this mask to extract the 3D ROI \( R \subset I \).

BFS-DFS Pipeline

Figure 1: BFS-DFS Pipeline for fetal brain segmentation.

In the DFS stage, Models B, C, and D progressively refine \( R \) through sliding-window passes, each stage narrowing the ROI and updating a binary mask \( S_{final} \) via majority voting across maps \( S_B, S_C, \) and \( S_D \), leading to an optimized fetal brain mask with minimized false positives.

Label Maps

Training involves generating synthetic shapes around the brain label maps, which represent human-annotated segmentations of the fetal brain. To create synthetic images from these label maps, we utilized the following parameters, indicating the standard deviation for each parameter:

Parameter Table

Table 1: Model parameters and sliding window configurations.

Here are two examples illusterating this augmentation where labels 1 to 7 point to brain structures and others belong to the background:

Synthetic Shapes Example

Figure 2: Example of synthetic shapes surrounding the brain label map to simulate diverse training conditions.

Training Data

Here are two examples of training data derived from the label maps: one for Model A, which is approapriate for the largest patches in the image, and another for Model D, which targets the smallest patches.
Synthetic Shapes Example

Figure 2: Example of synthetic images derived from the label maps for model \( A \) and model \( D \).

Key Advantages:

  • False Positive Reduction: BFS and DFS steps progressively refine the region of interest, minimizing false positives.
  • Adaptability to Sparse Data: Trained on synthetic data, the model performs well even with limited annotations.
  • Precision Across Variability: Multi-scale sliding windows capture diverse head orientations and anatomy for accurate segmentation.
  • High Accuracy: Outperforms existing methods, improving Dice scores by up to 5% on second-trimester and EPI scans.
  • Reliable Localization: The initial localization process ensures accurate brain segmentation, allowing for finer navigation within the womb and reducing the risk of segmenting non-relevant areas.

This repository is an invaluable resource for medical imaging problems, particularly for body organ segmentation. It provides a robust tool for addressing challenges related to limited training data and sparse annotations.


ECG2AF Broad Institute 2024/10/29 AI for AF Prediction

ECG2AF Model Web Application

This application demonstrates our ECG2AF tool, a clinical AI model designed to predict the risk of developing atrial fibrillation (AF) from ECG data.

We have shut down the online demo temporarily but to run this app locally, place app.py, ecg_model.py, and the required folders inside the app/ directory. Ensure that Dockerfile and requirements.txt are in the parent directory.

I developed this project using editors like vim and VS Code, and also Copilot for some function autocompletion. The app was (will be) deployed on an EC2 instance (Linux t2.medium), with an attached 32 GB gp3 volume for additional storage. Additionally, We’ve set the Nginx file upload limit to 500 MB to accommodate larger files.

Background

AI is widely used in clinical applications to improve risk stratification and intervention. This project focuses on cardiovascular disease, aiming to predict AF risk based on ECG data.

Objective

Our objective is to allow users to:

  • Upload an ECG file (.hd5 format)
  • Process the uploaded ECG using a pre-trained ECG2AF model
  • Display the prediction results, including four output values

How It Works

  1. Upload your file(s) in .hd5 format by dragging and dropping it into the upload area. Upload ECG File
  2. Process with ECG2AF: The model processes the file in real-time.
  3. View Predictions: The app displays four predictions for each ECG file, including AF risk and demographic estimations. You’ll also see charts for better interpretation. Example Results

Prediction Outputs

  • AF Risk: Estimated risk for developing atrial fibrillation.
  • Sex Prediction (Male/Female): Probability-based estimate of biological sex.
  • Age Prediction: Estimated age based on ECG data.
  • AF In Read (Yes/No): Classification for AF presence in the read.

Scalability Considerations

  • Batch Processing: Process multiple ECG files simultaneously.
  • Database and Caching: Store previously uploaded files separately for each user.
  • Cloud Deployment and Load Balancing: AWS’s Elastic Load Balancer (ELB) is used for additional scaling.

Visualization and Results

The charts in the application are created using Chart.js. Each chart presents predictions related to the uploaded .hd5 files, including AF risk, sex prediction, and age prediction values.

Example Results

Adding Nginx to Flask with Docker

Nginx acts as a reverse proxy for our Flask app in Docker, handling HTTP requests on port 80 and forwarding them to Flask on port 5000.

Benefits of Using Nginx

  • Security: Nginx hides the Flask server from direct internet exposure.
  • Load Balancing: Manages traffic efficiently and scales easily.
  • Static File Handling: Serves static files faster, reducing Flask's load.
  • Production-Ready: Handles multiple users reliably, unlike Flask's dev server.

If you have any questions, feel free to email me.


SEQ2SQL YINS/YALE 2024/10/29 Text to SQL

What’s the Scoop?

This project is all about turning natural language questions into SQL commands. We’ve got a model that uses LSTM and GRU units to make sense of your questions and turn them into SQL. Think of it as chatting with your database!

Datasets We Played With

We trained on some cool datasets:

  • ATIS: Questions about flights.
  • GEO880: Geographic info with 880 examples.
  • Overnight: Covers 11 fun domains like sports and restaurants.
  • Toy: A small dataset with 600 examples.
  • WikiSQL: A solid 1,000 examples for generating SQL queries.

Query Examples

Check out some killer examples:

What’s the description of a CH-47D Chinook?
SELECT col1 FROM table_1_10006830_1 WHERE col0='CH-47D Chinook';
What’s the max gross weight of the Robinson R-22?
SELECT col2 FROM table_1_10006830_1 WHERE col0='Robinson R-22';
What school did player number 6 come from?
SELECT col5 FROM table_1_10015132_1 WHERE col1='6';

How It Works

So, how does it all come together? The model learns to associate pairs of natural language questions with their corresponding SQL commands, determining the best SQL query based on the input question. I accomplished this during my first year of grad school at Yale! I’m thrilled to see that my repository has received 93 stars and 23 forks on GitHub so far.

If you have any questions, feel free to email me.


EM4QT University of Tehran 2016/10/29 CLIR

Dictionary-based Query Translation in CLIR

So, what’s the deal with this project? It’s all about improving the query language model by using top-ranked documents in a cross-lingual information retrieval (CLIR) setting. We use the Information derived from both source and target language documents, pulling them based on user queries.

EM4QT

Query Translation Idea

The main idea behind EM4QT is to grab relevant documents in both languages. When someone asks a question in their native language, we kick things off by retrieving a set of documents that match that query. This gives us our initial batch of source language documents.

However, at the same time, we whip up an initial guess of translation candidates using a bilingual dictionary.

Once we have both sets of documents, we fine-tune our translation candidates based on the top retrieved documents. This back-and-forth process makes sure we consider the most relevant and contextually appropriate translations, boosting the accuracy of our cross-lingual information retrieval.

Expectation-Maximization Algorithm

The EM algorithm is all about upping the translation quality by leveraging those top-retrieved documents in both languages. It boils down to two main steps:

  • E-Step: We estimate the translation confidence probabilities using the documents we’ve retrieved.
  • M-Step: Next, we maximize the likelihood of these probabilities, refining the translation model as we go.

This iterative refinement keeps going until we reach convergence, resulting in a more accurate translation distribution that enhances the overall quality of our cross-lingual retrieval. The full detail of the algorithm is published at IP&M Journal. If you have any questions, feel free to email me.


Education

Year Degree Field Institution Location
2023 PhD Computer Science Yale University USA
2015 MEng Software Engineering University of Tehran Iran

Research Interests

I am broadly interested in medical image analysis, machine learning, Church Doorspredictive modeling, and data-driven methods and precision medicine, with a focus on:.

- Brain Segmentation
- Fetal Brain Extraction
- Functional Connectomics and Brain-behavior Associations
- Optimal Transport and Data-driven Methods for Conectomics Data
- High-Performance Computing

Honors and Scholarships

Year Award Details
2022 Best Paper Award Graphs in Biomedical Imaging, MICCAI (url)
2022 MICCAI Student Award Early Acceptance, Singapore
2021 Brain Initiative Trainee Award Flash talk url
2020 Best Poster Award Connectomics for NeuroImaging, MICCAI (url)
2012 Graduate School Fellowship Ranked 5th in Undergraduate Studies, University of Tehran
2008 Ranked 331st/220,000 Top 0.2% in Iran's Nationwide University Entrance Exam