Sukhrob Ilyosbekov

Cell Painting images show how cells change after a drug or a genetic perturbation. MorphoCLIP learns to match those images to a plain-language description of the treatment, in either direction. The image and text encoders (DINOv3 and BioClinical ModernBERT) stay frozen and only small projection heads are trained, so it fits on one consumer GPU. We evaluated it on CPJUMP1, which has 51 plates, over 3 million images, 303 compounds, and 160 genes. Plate effects are corrected in the embedding space so the model can't score well just by recognizing which plate an image came from. I led the design, training, and evaluation, working with two co-authors at Northeastern.

@article{ilyosbekov2026morphoclip,
  title   = {MorphoCLIP: Text-Supervised Contrastive Learning for
             Perturbation Matching in Cell Painting Images},
  author  = {Ilyosbekov, Sukhrob and Gajjar, S. and Jin, R.},
  journal = {arXiv preprint arXiv:2608.22690},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.22690}
}

Ask a commercial image editor to change someone's nose and it will often retouch the whole face, which defeats the point of a surgical preview. This paper tests how much of that a client can fix without access to the model. I compared prompt-only steering, masked compositing, and model-based inpainting across six commercial editors and one inpainting model on 196 facelift and rhinoplasty edits. Identity preservation was measured with ArcFace, along with whether each edit stayed where it was asked to. The simplest method, masked compositing, kept edits in place better than inpainting did.

@article{ilyosbekov2026localize,
  title   = {Localize, Don't Beautify: Client-Side Control of
             Image-Editing APIs for Cosmetic Surgery Previews},
  author  = {Ilyosbekov, Sukhrob},
  journal = {arXiv preprint arXiv:2608.02841},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.02841}
}

A skin-lesion classifier covering all nine ISIC 2019 categories: EfficientNet V2 at 384x384, trained with focal loss because some classes are rare. It reaches about 86% accuracy (0.86 weighted F1) on roughly 25,000 dermoscopic images. The other half of the paper is about explanations. GradCAM++ attention is split along the ABCDE criteria dermatologists already use, and asymmetry, border irregularity, color variation (via K-means), and diameter are all measured from the lesion mask. Since those features are numbers, the paper can score how well the model's attention matches them instead of showing a few hand-picked heatmaps.

@article{ilyosbekov2025melanomanet,
  title   = {MelanomaNet: Explainable Deep Learning for
             Multi-Class Skin Lesion Classification},
  author  = {Ilyosbekov, Sukhrob},
  journal = {arXiv preprint arXiv:2512.09289},
  year    = {2025},
  url     = {https://arxiv.org/abs/2512.09289}
}