Research

Projects & open research

brain@lab:~$ git clone BRAIN-Lab-AI/
SegVLM — SegVLM architecture

SegVLM: Deformable Attentive Visual Enhancement for Referring Image Segmentation Using Vision-Language Model

Alaa Dalaq1, Muzammil Behzad1,2,*
1KFUPM · 2SDAIA-KFUPM JRC-AI

TL;DR

Give it an image and a phrase like “black baseball cap” and it returns a pixel-accurate mask — sharpening visual features with deformable convolutions and SE recalibration before they meet the language, for +10.3 mIoU over CRIS on PhraseCut.

UDAN-CLIP — UDAN-CLIP pipeline

UDAN-CLIP: Underwater Diffusion Attention Network with Contrastive Language-Image Joint Learning for Underwater Image Enhancement

Afrah Shaahid1, Muzammil Behzad1,2
1KFUPM · 2SDAIA-KFUPM JRC-AI

TL;DR

A diffusion model that restores color, contrast and fine detail in murky underwater photos, steered by a CLIP classifier that keeps natural in-air priors and a spatial attention module that targets haze and backscatter.

PULSE — framework overview

PULSE: A Unified Multi-Task Architecture for Cardiac Segmentation, Diagnosis, and Few-Shot Cross-Modality Clinical Adaptation

Hania Ghouse1, Maryam Alsharqi2, Farhad Nezami2,3, Muzammil Behzad1,4
1KFUPM · 2MIT · 3Harvard Medical School · 4SDAIA-KFUPM JRC-AI

TL;DR

One model reads a short-axis cardiac MRI study and returns ventricular segmentation, cardiomyopathy diagnosis and a structured clinical report in a single pass — and transfers to new scanners and even echocardiography.

KRCapVLM — KRCapVLM framework

KRCapVLM: Beam-Guided Knowledge Replay for Knowledge-Rich Image Captioning using Vision-Language Model

Reem Aljunaid1,2, Qasim Umer1, Sajjad Mahmood1, Mahmood Niazi1, Muzammil Behzad1,3,*
1KFUPM · 2Imam Abdulrahman Bin Faisal University · 3SDAIA-KFUPM JRC-AI

TL;DR

Captioning models say “a tower”; KRCapVLM says “Big Ben”. Beam-guided knowledge replay, patch self-attention and a cosine schedule lift real-world concept recognition from 50.4% to 63.3% on KnowCap.

Attack-Aware — dual-stream architecture

Attack-Aware Deepfake Detection under Counter-Forensic Manipulations

Noor Fatima1, Hasan Faraz Khan1, Muzammil Behzad1,2,*
1KFUPM · 2SDAIA-KFUPM JRC-AI

TL;DR

A deepfake detector built to survive the edit: it trains on whichever compression, regrain or transcode attack hurts most, fuses semantic and forensic-residual streams, and shows where the evidence is with weakly supervised heatmaps.

MultiviewVLM — MultiviewVLM framework

MultiviewVLM: Unsupervised Multiview Contrastive Language-Image Joint Learning with Pseudo-Labeled Prompts via Vision-Language Model for 3D/4D Facial Expression Recognition

Muzammil Behzad1,2
1KFUPM · 2SDAIA-KFUPM JRC-AI

TL;DR

A vision-language model that learns facial-emotion representations from 3D/4D faces without labels — text-prompt pseudo-labels supply the emotional meaning, and multiview contrastive learning ties together different views of the same face.

Collaborate

Have data, compute or a hard problem?

We partner with hospitals, industry and academic groups on applied AI research. Let's talk about co-supervised projects, datasets and joint grants.