PhD Researcher · Albert Einstein College of Medicine
Working where computation meets cancer biology.
I study how metastatic cancer re-disseminates — cells that don't follow a straight line.
Neither has my path: biomedical engineering, code, AI, and now the wet lab. This is a record of that route.
01
Two fields, one bench
The biology
Cancer metastasis & re-dissemination
Understanding how tumor cells spread, seed new sites, and re-disseminate — the biology behind why metastasis is so hard to stop. Now expanding into experimental wet-lab work.
Metastasis
Tumor biology
Wet lab
Histology
The computation
AI & computational biology
Building models and pipelines to make sense of complex biological data — bringing an engineer's toolkit to questions in oncology.
Machine learning
Python
Comp. biology
Data pipelines
Current focus
What makes metastatic cancer come back?
My doctoral research investigates metastatic re-dissemination — how cancer that has already spread can seed further sites. The aim is translational: insight that could change how recurrence is understood and treated.
Funded by a METAvivor Translational Research Award
02
The path here wasn't a straight line
2025 — now
PhD, Biomedical Research
Albert Einstein College of Medicine · New York
Cancer metastasis re-dissemination. Deliberately expanding from computation into experimental biology.
Prior
M.S.
Johns Hopkins University
[Add your degree focus and a line on what you worked on.]
Earlier
Biomedical Engineering · Programming · AI
[Institution / roles]
[The foundation — the "detour" that turned out to be the point.]
03
Selected work
Project 01 · Publication · npj Breast Cancer (2026)
Racial differences in TMEM doorways predict breast cancer recurrence
Black women with ER+/HER2− breast cancer experience higher rates of recurrence and mortality than
White women, even after accounting for differences in access to care and standard clinical factors.
This study investigated whether those disparities are reflected in the tumor microenvironment
before treatment begins.
418
Treatment-naïve samples
4
NYC institutions
4.6×
5-yr recurrence risk
2026
npj Breast Cancer 12:3
What we did
We quantified TMEM doorways — tri-cellular structures composed of a tumor cell, a
Tie2-high macrophage, and an endothelial cell that serve as portals for tumor cell
intravasation — in 418 treatment-naïve breast cancer specimens collected from four
New York City institutions.
Key findings
Black patients had significantly higher TMEM doorway scores than White patients in ER+/HER2−
and HER2+ breast cancers, whereas no significant difference was observed in triple-negative
breast cancer.
Among patients with ER+/HER2− disease, elevated TMEM doorway scores were associated with a
4.6-fold higher risk of recurrence within five years for Black patients, while no
significant association was observed for White patients. This interaction remained significant
after adjusting for tumor size, grade, nodal status, and treatment.
My contribution
Applied a computational pathology pipeline for automated TMEM doorway identification.
Imported, organized, and managed multi-institutional whole-slide imaging datasets.
Developed tissue segmentation using a Bayesian classifier in Visiopharm to identify tumor cells, macrophages, and endothelial cells.
Implemented morphometric image analysis algorithms to detect TMEM doorways in panMena/CD68/CD31 triple-immunohistochemistry slides.
Performed statistical analyses to quantify baseline differences in the tumor microenvironment across racial groups.
Identified significantly higher baseline TMEM doorway densities and macrophage infiltration in Black patients with ER+/HER2− and HER2+ breast cancers.
Digital pathology
Visiopharm
Bayesian classification
Image analysis
Whole-slide imaging
Logistic regression
R
Figures
assets/figures/fig1-consort.png
Figure 1. Patient cohort and study workflow. 418 treatment-naïve invasive ductal
carcinomas from four NYPOG institutions, stained for panMena, CD68, and CD31 and grouped
by self-identified race and receptor subtype.
assets/figures/fig2-tme-by-race.png
Figure 2. Baseline differences in TMEM doorway score and tumor microenvironment
parameters by race. Differences appear in ER+/HER2− and HER2+ tumors, but not in TNBC.
Microvascular density does not differ by race in any subtype.
assets/figures/fig3-forest-interaction.png
Figure 3. Interaction between TMEM doorway score and race in predicting recurrence
risk, adjusted for age, BMI, tumor size, nodal status, grade, and treatment.
Parmar P, Karadal-Ferrena B, Shukla S, et al. Racial disparity in pro-metastatic tumor
microenvironment in treatment-naïve breast cancer. npj Breast Cancer 12, 3 (2026).
Figures reproduced from the original article under
CC BY 4.0.
Project 02 · Code · Zenodo (2026)
An image analysis pipeline for quantifying single-, double-, and triple-positive cells in breast cancer tissue
Off-the-shelf tools can't reliably segment IF-stained tissue into individual cells, track which
fluorescent signal belongs to which cell, and classify each one's phenotype — so I built an
end-to-end pipeline that does all three.
4
IF channels processed
0.987
Pearson r vs. manual counts
3
Phenotype classes (1×/2×/3×)
10%
Stroma-overlap QC cutoff
The problem
Most breast cancer deaths come from metastasis, and a key driver is cancer cells intravasating
through TMEM (Tumor MicroEnvironment of Metastasis) doorways. The cells that matter most
are those co-expressing dormant (NR2F1), invasive (MenaINV), and stem-like
(SOX9) markers — but quantifying them at scale is a genuine bottleneck.
Off-the-shelf tools can't reliably segment IF-stained tissue into individual cells and separate
their nuclear and cytoplasmic compartments while keeping track of which signal belongs to which
cell.
What I built
Nuclear segmentation using StarDist, a deep-learning model suited to densely packed nuclei.
Whole-cell inference via Voronoi tessellation to assign cytoplasmic territory around each nucleus.
Stroma removal by thresholding the MenaINV channel and applying iterative erosion/hole-fill/dilation, then subtracting the resulting mask from the nuclear, cell-body, and cytoplasmic masks.
Quality control excluding cells with >10% stromal overlap or touching the image boundary.
Feature extraction of mean intensity, area, and spatial coordinates per compartment.
Hierarchical phenotyping into single-, double-, and triple-positive cells using compartment-specific thresholds (negative control + quartile/tertile/decile cutoffs).
Spatial analysis including nearest-neighbor interactions and hotspot detection.
Result & why it matters
Automated counts matched manual expert scoring almost perfectly (Pearson r = 0.987,
p = 0.00026), with nuclear/cytoplasmic segmentation verified by visual inspection.
The pipeline links cell phenotype to spatial context, making it possible to map how the most
aggressive tumor cells organize around TMEM doorways — the next step for this work, and a route
into mechanisms of metastasis and therapy response.
ImageJ / Fiji
StarDist
Voronoi tessellation
Image analysis
Spatial statistics
My role
Developed the segmentation and compartment-isolation pipeline end to end.
Designed the phenotype classification logic and ran threshold-sensitivity testing.
Ran all downstream statistical analysis, including the validation benchmark against manual counts.
Figures
assets/figures/fig4-tpp-pipeline.png
Figure 1. Pipeline schematic, from raw multiplex IF images to per-cell phenotype calls.
assets/figures/fig5-tpp-validation.png
Figure 2. Validation scatter plot: automated vs. manual cell counts (Pearson r = 0.987, p = 0.00026).
assets/figures/fig6-tpp-masks.png
Figure 3. Compartment mask processing — nuclear, cell-body, cytoplasmic, and stroma masks per field of view.
assets/figures/fig1-consort.png
assets/figures/fig2-tme-by-race.png
assets/figures/fig3-forest-interaction.png
