Multimedia Interaction and Communication Lab AAST

𝗡𝗲𝘄 𝗣𝗿𝗲𝗽𝗿𝗶𝗻𝘁 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗠𝗜𝗖 𝗟𝗮𝗯: Meezan, An AI-Assisted Coaching System for Greco-Roman Wrestling Using a Norm-Referenced, Factor-Based Elite-Distance Framework

𝗠𝗼𝘁𝗶𝘃𝗮𝘁𝗶𝗼𝗻: Coaches need better tools to gauge how far a given wrestler's physical and technical skill profile diverges from an elite reference standard.

𝗠𝗲𝘁𝗵𝗼𝗱𝗼𝗹𝗼𝗴𝘆: Explored a validated 31-item technical and physical test battery with an elite group of 91 Greco-Roman wrestlers (ages 13-17). Used Exploratory Factor Analysis (EFA) to discover latent factor structure. We then developed a norm-referenced, factor-based elite-distance framework, using a robust-covariance Mahalanobis distance to quantify overall profile atypicality. Integrated with a Large Language Model (LLM) to generate personalized coaching reports.

𝗣𝗿𝗲𝗺𝗶𝘀𝗲: AI-driven individualized assessment and coaching system that goes beyond predictive framings to characterize how an athlete deviates from the typical elite profile across specific latent factors.

Check out the full preprint for details!

https://doi.org/10.51224/SportRxiv.1008

We are pleased to share our latest preprint: "Phenotyping Tibial Plateau Fractures via Self-Supervised Learning: A Label-Agnostic Framework with Expert Validation."

Tibial plateau fracture classification relies on schemes like Schatzker and AO/OTA — but these suffer from well-documented inter-observer variability. Supervised AI trained on inconsistent labels learns human disagreement, not fracture morphology.

Our work takes a different approach entirely. We train a RadImageNet-pretrained ResNet-50 encoder using SimCLR contrastive learning — with no labels — to discover "N" stable, imaging-derived fracture phenotypes directly from knee radiographs.

Key results:
1. Bootstrap ARI = 0.319 ± 0.041 — stable clusters across 100 resampling iterations
2. Silhouette = 0.511 — strong internal geometric cohesion
3. Inter-partition ARI vs Schatzker = 0.013 — 𝗽𝗵𝗲𝗻𝗼𝘁𝘆𝗽𝗲𝘀 𝗮𝗿𝗲 𝗼𝗿𝘁𝗵𝗼𝗴𝗼𝗻𝗮𝗹 𝘁𝗼 𝗰𝗼𝗻𝘃𝗲𝗻𝘁𝗶𝗼𝗻𝗮𝗹 𝗹𝗮𝗯𝗲𝗹𝘀, 𝗯𝘆 𝗱𝗲𝘀𝗶𝗴𝗻
4. Blinded expert validation by two independent orthopaedic clinicians — all four phenotypes met the pre-specified face validity threshold
5. One phenotype was unanimously identified as capturing comminution and high-complexity injuries — without a single label seen during training

𝗣𝗲𝗿𝗵𝗮𝗽𝘀 𝘁𝗵𝗲 𝗺𝗼𝘀𝘁 𝘁𝗵𝗼𝘂𝗴𝗵𝘁-𝗽𝗿𝗼𝘃𝗼𝗸𝗶𝗻𝗴 𝗳𝗶𝗻𝗱𝗶𝗻𝗴: expert reviewers anchored to Schatzker vocabulary perceived our groups as heterogeneous precisely where Schatzker alignment was lowest — suggesting that clinician perception and label-agnostic geometry are measuring orthogonal dimensions of the same morphological space. We argue this is not a limitation of the method. It is the finding.

Preprint available on arXiv at the following link:
http://arxiv.org/abs/2606.17295

Feedback and collaboration inquiries are very welcome.

We are pleased to share that our new preprint is now live on arXiv:

📄 3D-CBM: A Framework for Concept-Based Interpretability in Generative 3D Modeling

🔗 https://lnkd.in/duPpcgjt

Modern 3D generative models can produce stunning geometry — but they remain largely opaque. A designer cannot tell a model that a chair leg is too thin to bear load, or that a joint is anatomically incorrect, because there is no shared conceptual vocabulary between human intent and model internals.

This work proposes 3D-CBM: a framework that embeds Concept Bottleneck Models directly into a 3D generative architecture, creating a "glass-box" layer of human-interpretable concepts between raw point cloud input and reconstructed output.

Key contributions:

→ A formal mathematical framework extending CBMs to unstructured 3D data (point clouds & meshes)

→ A three-tier concept taxonomy: geometric primitives → structural components → functional attributes

→ A standardized Human-in-the-Loop intervention protocol for test-time concept overriding

→ A proof-of-concept experiment on PartNet chairs achieving 88.8% concept accuracy and a Chamfer Distance of 0.0115

→ A re-encoding validation methodology that closes the HITL loop and rigorously confirms semantic fidelity — going beyond geometric perturbation metrics alone

The intervention analysis reveals an important asymmetry: concept suppression (e.g., removing legs) is highly effective, while concept imposition (e.g., forcing armrests) is counteracted by the unchanged latent embedding — a finding that motivates intervention-aware training as a concrete direction for future work.

This is, to our knowledge, the first work to formally incorporate CBMs into a 3D generative pipeline, filling a gap explicitly identified across the recent interpretability literature.

This research is sponsored by Yubree Labs - formerly VRapeutic.

We are pleased to announce our latest preprint, introducing PAIRED — our proposed framework for transparent reporting of AI contributions in scientific research.

In this work, we define a process-anchored approach for attributing AI contributions, something that is entirely absent from current disclosure frameworks. Built around a four-stage research pipeline, PAIRED defines five principal roles for AI arranged along an agency gradient, a unit of attribution called the decision point, and an artifact-trigger rule that governs when a three-field micro-log entry must be recorded.

We demonstrate PAIRED through three worked examples drawn from a recent preprint from our lab, and we close with an Acknowledgments section that is itself a meta-demonstration of PAIRED — documenting the very article that proposes it using the framework's own micro-log format.

As part of the discussion, we propose a concrete adoption pathway and issue a call-to-action to AI platform developers: integrating PAIRED as an opt-in feature in language models that offer a Deep Research functionality, so that the model assists with bookkeeping while the researcher retains full epistemic authority over the final record.

The preprint is available at: https://arxiv.org/abs/2605.24325

Excited to share our latest preprint on AI-powered swimming analytics and coaching support systems.

In this paper, we propose a novel Retrieval-Augmented Generation (RAG) framework that combines multimodal physiological data, IMU sensing, domain literature, and expert coaching knowledge to synthesize high-quality datasets for trustworthy AI applications in swimming science.

Our work explores how Large Language Models (LLMs) and synthetic data generation can help overcome challenges such as limited aquatic datasets, privacy constraints, and expensive expert labeling—while enabling more context-aware and technically grounded AI support for athletes and coaches.

The study also introduces a multi-agent LLM architecture for generating and validating expert-level “Question–Context–Answer” triplets aligned with evidence-based sports science principles.

We hope this work contributes toward more reliable, explainable, and practically deployable AI systems in sports performance analysis and coaching.

The preprint can be found through the following link:

https://lnkd.in/dHN439de

🚀 Excited to share our latest work, now available as a preprint!

📄 "Do Hybrid CNN–Transformer Architectures Really Generalize? A Systematic Review of Cross-Organ, Multi-Modal, and Multi-Dataset Transfer in Medical Imaging"

As AI moves closer to real-world clinical deployment, generalization — not just performance — becomes the defining challenge. In this systematic review, we synthesize evidence from 71 peer-reviewed studies (2020–2025) to critically examine whether hybrid CNN–Transformer models truly deliver on their generalization promise in medical imaging.

🔍 What we cover:
→ A structural taxonomy of 5 recurring hybrid architectural patterns
→ Generalization across cross-organ, cross-dataset, and multi-modal scenarios
→ A critical assessment of reproducibility and evaluation rigor in the field
→ A forward-looking roadmap for efficiency-aware, clinically deployable AI

💡 Key takeaway: Hybrid models show strong promise, but high computational cost, limited external validation, and over-reliance on fully supervised learning remain significant barriers to clinical translation.

https://lnkd.in/dNWAfEJV

We’re hiring: 1 Part-Time Research Assistant for Research on CNN-Transformer Hybrid Models (3 months, extendable) — MIC Lab (Remote-friendly).

The Multimedia Interaction and Communication (MIC) Lab is opening one part-time RA position for a 3-month research project (extendable) starting immediately. If you’re a fresh graduate eager to build serious, hands-on research experience in medical imaging & AI, we’d love to hear from you.

✅ MIC Lab provides computational resources to support your experiments and training runs.

Must-haves
— Fresh graduate (Bachelor’s or Master’s; Class of 2023–2025).
— Background in computer vision / medical imaging / machine learning.
— Strong proficiency with deep learning frameworks (e.g., PyTorch / TensorFlow).
— Strong reading & synthesis skills (papers → structured insights).
— Excellent academic writing in English.
— Reliable part-time commitment for 3 months; ability to deliver weekly progress.

— Strong experience in CNN-transformer hybrid models AI for classification and/or segmentation, with clear evidence through projects, code, or research outputs.

Nice-to-haves
— Proven experience writing research papers (drafting, revising, responding to feedback).
— Experience with LaTeX and reference managers.

Who should apply
— Fresh graduates seeking rigorous, real-world research experience to launch a career in CV/AI.
— Detail-oriented self-starters who can own tasks end-to-end (reading → experiments → write-ups).

How to apply
Email [email protected] with subject line:
“Part-Time Research Assistant – MIC Lab – Spring 2026”

Include:
1. Your resume
2. GitHub / Google Scholar / portfolio (if available)
3. A link to one recent paper you liked (relevant to the scope of the RA opening) + 1–2 sentences on what it adds

Start date: Immediately (3-month part-time commitment — extendable)
Remote/Location: Remote-friendly; meeting times aligned with GMT +2 / +3
=====
About MIC Lab:
MIC Lab is a research group, a part of Ethri Labs, and spin-off the Arab Academy for Science and Technology (AAST), Egypt. We conduct research on multimedia systems, focusing on n-dimensional signal coding/compression and synthesis, innovative interaction modalities, and HCI-grounded evaluation of intelligent systems and experiences.

If you know someone who fits, please share or tag them 👇

Bridging the Gap Between AI Accuracy and Clinical Trust in Neurovascular Care

Artificial Intelligence is transformative for neuroimaging, but the "black-box" nature of deep learning remains a critical barrier to surgical adoption. If a model flags an intracranial aneurysm as high-risk, a surgeon needs to know why.

In our latest preprint, we introduce an end-to-end 3D Concept Bottleneck Model (CBM) framework that brings human-understandable reasoning to aneurysm classification.

The Impact: Unlike traditional AI that provides post-hoc visual heatmaps, our framework constrains the model’s internal logic to 30 real-world clinical indices—such as vessel geometry and hemodynamics.

Key Results:
- High Performance: Achieved peak diagnostic accuracies of 93.33% (ResNet-34) and 91.43% (DenseNet-121).
- Reliability: Maintained an accuracy-generalization gap of < 0.04, ensuring stable results across diverse clinical data.

Trust by Design: By anchoring AI predictions to neurosurgical principles, we provide clinicians with a verifiable "reasoning" path for better planning.

Worldwide Potential: By prioritizing clinical transparency alongside predictive power, this framework offers a scalable solution for healthcare systems worldwide to adopt AI-assisted diagnostics that physicians can actually trust.

Here is the link to the preprint:
https://lnkd.in/dYaT3hWV

We’re hiring: 2 Part-Time Research Assistants (3 months, extendable) — MIC Lab (Remote-friendly).

The Multimedia Interaction and Communication (MIC) Lab is opening two part-time RA positions for a 3-month research project (extendable) starting immediately. If you’re a fresh graduate eager to build serious, hands-on research experience in medical imaging & AI, we’d love to hear from you.

✅ MIC Lab provides computational resources to support your experiments and training runs.

Must-haves
— Fresh graduate (Bachelor’s or Master’s; Class of 2023–2025).
— Background in computer vision / medical imaging / machine learning.
— Strong proficiency with deep learning frameworks (e.g., PyTorch / TensorFlow).
— Strong reading & synthesis skills (papers → structured insights).
— Excellent academic writing in English.
— Reliable part-time commitment for 3 months; ability to deliver weekly progress.

Two openings (choose your best fit)
1) Generative Modeling Track
— Strong, demonstrable experience with GANs / generative modeling, ideally evidenced by projects, GitHub repos, or publications.

2) Explainability Track (Classification/Segmentation)
— Strong experience in explainable AI for classification and/or segmentation, with clear evidence through projects, code, or research outputs.

Nice-to-haves
— Proven experience writing research papers (drafting, revising, responding to feedback).
— Experience with LaTeX and reference managers.

Who should apply
— Fresh graduates seeking rigorous, real-world research experience to launch a career in CV/AI.
— Detail-oriented self-starters who can own tasks end-to-end (reading → experiments → write-ups).

How to apply
Email [email protected] with subject line:
“Part-Time Research Assistant – MIC Lab – Fall 2025”

Include:
1. Your resume
2. GitHub / Google Scholar / portfolio (if available)
3. A link to one recent paper you liked (relevant to the track you’re applying for) + 1–2 sentences on what it adds

Start date: Immediately (3-month part-time commitment — extendable)
Remote/Location: Remote-friendly; meeting times aligned with GMT +2 / +3
=====
About MIC Lab:

MIC Lab is a research group at the Arab Academy for Science and Technology (AAST), Egypt. We conduct research on multimedia systems, focusing on n-dimensional signal coding/compression and synthesis, innovative interaction modalities, and HCI-grounded evaluation of intelligent systems and experiences.

If you know someone who fits, please share or tag them 👇

Domain shift can make top-performing medical imaging models fail catastrophically on external datasets, hindering trustworthy clinical AI and motivating Explainable AI (XAI) as a diagnostic tool. We present a rigorous two-phase diagnosis of UMamaba (a state-space model) for cerebrovascular segmentation. First, we quantify a domain gap between RSNA CTA Aneurysm (Source) and TopCoW Circle of Willis CT (Target), driven by differences in Z-resolution and background noise, and observe a sharp performance drop. Second, our core contribution uses Seg-XRes-CAM to quantify attention overlap with Ground Truth and predictions, showing attention abandons true anatomy and follows spurious cues. 0.8604 → 0.2902; IoU ≈ 0.101 (GT), IoU ≈ 0.282 (pred), at 0.3 threshold.

Check the preprint through the following link:
http://arxiv.org/abs/2512.13977