>

PORTRAIT, PLEASE

This site is built for portrait.

Hello.

MINA
HEINEIN.

AI & Digital Health engineer. I work on making biomedical AI reliable and transparent.

Looking for M2 research internships and part-time work

Portrait of Mina Heinein
Collège de la Salle logo

EDUCATION 01 · COLLÈGE DE LA SALLE, CAIRO

A French-language education.

I completed my primary and secondary education at Collège de la Salle in Cairo, where French was the language of instruction. It is the source of my working French (B2), and part of why Paris is the next step.

Mina as a schoolboy in uniform
CAIRO
Mina at his school graduation
GRADUATION DAY
Cairo University crestDFKI logo

EDUCATION 02 · CAIRO UNIVERSITY · 2021–2026

B.Sc. Systems & Biomedical Engineering.

Five years across biomedical engineering and computer science, with coursework and projects in medical imaging, signal processing, and embedded systems.

My graduation thesis, MedHyperGraph, was carried out in collaboration with DFKI in Germany: a temporally indexed structure linking a patient’s records, reports, and imaging so a model can reason across them. It was accepted as an oral at MICCAI 2026.

SEE THE WORK ↓
THE OBELISK & THE DOME · TRACED FROM CAMPUS · 30.026°N 31.211°E
PARIS, TRACED · 48.856°N 2.352°E
Université PSL logo

EDUCATION 03 · M.SC. (M2) AI & DIGITAL HEALTH · 2026–2027

NEXT · SEPT 2026

The next step: Paris.

The M2 in AI and Digital Health — ST4Health — at Université PSL runs 2026–2027. I move to Paris in September, and I am looking for an M2 research internship and part-time engineering work alongside the programme.

RESEARCH & ENGINEERING

The work.

I research medical-imaging AI and build the software around it. That work has produced two papers, both accepted as orals at MICCAI 2026, plus the engineering projects below.

News

  • AUG 2026Both papers accepted as orals at the MICCAI 2026 MSB EMERGE workshop, Strasbourg. Published on OpenReview 3 August.
  • SEP 2026Starting the M.Sc. (M2) in AI and Digital Health (ST4Health) at Université PSL, Paris.
  • 2026Graduated B.Sc. in Systems & Biomedical Engineering, Cairo University.

Research · Publications

Both papers were accepted as orals at the EMERGE workshop of MICCAI 2026, run by the MICCAI Student Board (MSB). Strasbourg, 27 September 2026.

ACCEPTED · MICCAI 2026 · MSB EMERGE WORKSHOP · ORALGRADUATION THESISFIRST AUTHOR · WITH DFKI

MedHyperGraph: EHR-Integrated Multimodal Hyperedges for Clinical VQA

M. Heinein*, M. Youssef*, K. Nashed*, T. Basha, H. M. T. Alam, A. M. Selim, O. S. Bhatti, D. Sonntag

*Equal contribution · with DFKI · funded by ASRT & ITIDA

ABSTRACT & RESULTS
PROBLEM

Clinical visual question answering should read the medical image and the patient’s record together, but most models only see the pixels.

APPROACH

Each case becomes a multimodal hypergraph that ties EHR entities to image findings. GraphRAG retrieves over that graph, so a vision-language model answers grounded in the real record, with no task-specific training.

RESULT

73.95% on EHRXQA boolean questions (n=1,900; 95% CI 71.9–75.9) without task-specific training — the only method that improves on all three backbones, and ahead of the 70.43% an RL-trained system reports on its own backbone. On open-ended multi-study questions weighted-F1 climbs from 5.2 to 43.4; ablations put most of the gain on the structured EHR.

APPROACH · HOW IT WORKS
EHR + medical imageMultimodal hypergraphGraphRAG retrievalVision-language modelGrounded clinical answer
A QUESTION FROM THE DEMO

“Does this image show cardiomegaly, and is there any abnormality in the left lung?”

The model pulls the patient’s prior studies and reports from the EHR hypergraph, reads the image, and answers using both sources together.

WITHDFKI, German Research Center for AI· FUNDED BYASRT, Academy of Scientific Research and TechnologyITIDA

73.95%

ACCURACY ON EHRXQA (n=1,900)

vs 70.43%

RL-TRAINED STATE OF THE ART · WITHOUT TASK-SPECIFIC TRAINING

5.2 43.4

WEIGHTED-F1 ON OPEN-ENDED · MULTI-STUDY JUMP

PythonPyTorchKnowledge GraphsHypergraphsGraphRAGVLMsNeo4j
BIBTEX
@inproceedings{heinein2026medhypergraph,
  title     = {MedHyperGraph: EHR-Integrated Multimodal Hyperedges for Clinical VQA},
  author    = {Heinein, M. and Youssef, M. and Nashed, K. and Basha, T. and
               Alam, H. M. T. and Selim, A. M. and Bhatti, O. S. and Sonntag, D.},
  booktitle = {MICCAI 2026 MSB EMERGE Workshop},
  year      = {2026},
  note      = {Oral presentation. Equal contribution: M. Heinein, M. Youssef, K. Nashed},
  url       = {https://openreview.net/forum?id=Gm7QeQBZxm}
}
ACCEPTED · MICCAI 2026 · MSB EMERGE WORKSHOP · ORALSECOND AUTHOR

LocSAM3: Box-Supervised Adaptation of SAM3 for Text-Only Chest X-Ray Segmentation

K. Nashed*, M. Heinein*, M. Youssef*, T. Basha, H. M. T. Alam, A. M. Selim, O. S. Bhatti, D. Sonntag

*Equal contribution · with DFKI

ABSTRACT & RESULTS
PROBLEM

Pixel-level masks are expensive to annotate on chest X-rays, where anatomy overlaps and pathological boundaries are ill-defined. Bounding boxes are far cheaper, and already exist in several datasets.

APPROACH

Adapts SAM3’s concept-grounding and localization using only boxes and concept names — no pixel masks. The pretrained mask decoder and over 95% of the vision backbone stay frozen, and the box is supplied for only a random subset of training samples, so the model learns to localize from text alone.

RESULT

On a 10-class MIMIC-CXR benchmark, text-only mIoU improves over Medical-SAM3 from 0.48 to 0.69 for anatomical structures and from 0.09 to 0.53 for pathological findings — with no dense masks in training and no spatial prompt at inference.

0.69

ANATOMY mIoU · TEXT-ONLY, FROM 0.48

0.53

PATHOLOGY mIoU · TEXT-ONLY, FROM 0.09

>95%

OF THE VISION BACKBONE KEPT FROZEN

PythonPyTorchSAM3SegmentationChest X-rayPEFT
BIBTEX
@inproceedings{nashed2026locsam3,
  title     = {LocSAM3: Box-Supervised Adaptation of SAM3 for Text-Only
               Chest X-Ray Segmentation},
  author    = {Nashed, K. and Heinein, M. and Youssef, M. and Basha, T. and
               Alam, H. M. T. and Selim, A. M. and Bhatti, O. S. and Sonntag, D.},
  booktitle = {MICCAI 2026 MSB EMERGE Workshop},
  year      = {2026},
  note      = {Oral presentation. Equal contribution: K. Nashed, M. Heinein, M. Youssef},
  url       = {https://openreview.net/forum?id=mWnH2hNROa}
}

Engineering · Selected builds

SignalSuite Toolkit

A digital signal processing suite: a real-time ECG/EEG viewer, an audio equalizer with vocal extraction and Wiener filtering, and a Shazam-style audio fingerprinter.

PythonPyQt5SciPy
GITHUB ↗
FIVE TOOLS INSIDE
  • Multi-Port Signal Viewer — real-time ECG/EMG/EEG streaming with synchronized graphs, cine playback with speed control, signal merging, and automated PDF reporting.
  • Signal Equalizer — slider-driven frequency editing with Uniform, Music & Animals, Vocal, and Wiener modes; live Fourier view on linear and audiogram scales, plus dynamic spectrograms.
  • Image Mixer — a Fourier magnitude/phase mixer that shows how much of an image each component carries, with adjustable weights and region selection.
  • Sampling Studio — sampling and reconstruction against the Nyquist rate, with configurable SNR to show how noise and aliasing interact.
  • Shazam-like App — audio fingerprinting via spectrogram + perceptual hashing, with ranked similarity matching across a track repository.

BUILT WITH A TEAM OF FIVE OVER A DSP COURSE SEQUENCE

Eduokee

Language learning through song lyrics, built on the Spotify API. It scores pronunciation in real time with Google Cloud Speech and Levenshtein distance, and lifted engagement by 45%.

ReactNodeSpotify API
GITHUB ↗
DEMO & SCREENSHOTS
  • Spotify Web API integration pulls the user’s saved playlists and liked songs, with 30-second track playback.
  • Lyrics are fetched per track, and every word is clickable for a definition and pronunciation.
  • Singing is recorded and transcribed with the Google Cloud Speech API.
  • Levenshtein distance compares the transcript against the real lyrics and scores the attempt out of 100.
  • Responsive for desktop and mobile.
Eduokee home page
HOME
Eduokee music player with lyrics
PLAYER · LIVE LYRICS
Eduokee pronunciation scoring results
PRONUNCIATION SCORING

Puppets.eg

A full-stack e-commerce store with custom orders and an admin dashboard. Conversion rose 25% and order-processing time dropped 60%.

ReactExpressMongoDB
GITHUB ↗
DEMO & FEATURES
  • React + Tailwind storefront with product listings, detail views, cart, and checkout.
  • A custom-order form lets buyers submit their own puppet request with an image upload.
  • Admin dashboard for managing products, orders, and incoming special requests.
  • Express REST API over MongoDB — /record for products, /orders, and /specialRequests.
  • End-to-end tests with Cypress.

AVR Smart Home

Bare-metal ATmega32 drivers with an ESP32 for Wi-Fi, so appliances can be controlled from the web in real time.

Embedded CATmega32ESP32
GITHUB ↗
SIMULATION & INTERFACE
  • Bare-metal ATmega32 drivers written from scratch: LCD, keypad, DIO, external interrupts, and UART.
  • An ESP32 bridges the microcontroller to a web server, relaying commands from the browser to the MCU.
  • Lights, locks, and appliances respond in real time from the web interface.
  • The whole system was simulated and tested in Proteus before it went to hardware.
Proteus circuit simulation of the smart home system
PROTEUS SIMULATION
Smart home web interface, main view
WEB INTERFACE
Smart home web interface, door lock control
LOCK CONTROL

Computer Vision Toolkit

Classical computer vision implemented from scratch: filtering, Canny and active-contour edge detection, SIFT matching, K-means and Otsu segmentation, and face recognition.

PythonOpenCV
REPO NOT PUBLIC
WHAT IT DOES
  • Filtering and noise models, then edge detection with Canny and active contours (snakes).
  • SIFT keypoint detection and descriptor matching.
  • Segmentation with K-means and Otsu thresholding.
  • A face recognition pipeline built on top of the same primitives.
  • Everything implemented from first principles rather than called from a library.

Chess Game

A full rules engine, including castling, promotion and timers.

JavaSwing
GITHUB ↗
WHAT IT DOES
  • A complete rules engine in Java: legal move generation, check and checkmate detection, castling, en-passant, and promotion.
  • Swing GUI with per-player timers.
  • Written as the object-oriented programming final project — the piece hierarchy and board state are the design exercise.

Internships

Mina at the EJADA office
EJADAAUG–SEP 2025

SOFTWARE ENGINEERING INTERN

Built CRM solutions on Microsoft Dynamics 365 for banking and government clients across the Gulf: REST/SOAP integrations, custom .NET plugins, and SLA-automated case management. My first close look at how regulation and procurement shape what actually ships.

Mina at the NAID internship
NAIDJUL–SEP 2024

WEB DEVELOPMENT INTERN

Built accessible web applications and assistive technology for users with mobility, sensory, and cognitive impairments. The same user-centred approach carries into clinical AI, where the users are clinicians and the stakes are higher.

TOOLING

Stack.

AI / ML

PyTorchRAG / GraphRAGVLMsSNOMED CT ontologies

SOFTWARE

CPythonJavaJavaScriptReactNodeNestJSFlask.NET

DATA

MongoDBPostgreSQLMySQLNeo4j

EMBEDDED

AVRSTM32

SPOKEN · ENGLISH (IELTS 8.0) · FRANÇAIS (B2) · العربية (NATIVE)

CREDENTIALS

Certificates.

SCROLL →

AUG 2024

ITI · INFORMATION TECHNOLOGY INSTITUTE

Embedded Systems Diploma

Covered microcontroller interfacing, real-time operating systems, embedded C, and automotive bus technology.

Embedded CRTOSAVR
NOV 2024

ALX AFRICA

Software Engineering Programme, Front-End Specialization

A 12-month project-based curriculum covering data structures and algorithms, DevOps, Nginx, load balancing, and technical interviewing.

JavaScriptDevOpsNginx
CERTIFIED

IELTS ACADEMIC

English, Band 8.0

Full professional proficiency in English, alongside French (B2) and native Arabic.

English 8.0Français B2العربية

Working on AI for healthcare? Let’s talk.

Mina's signature
© 2026 MINA HEINEIN.CAIRO → PARISUPDATED AUG 2026