Sukhrob Ilyosbekov

MorphoCLIP matches Cell Painting microscopy images of drug- and gene-perturbed cells to descriptions of the treatment written in ordinary language. The vision and language backbones stay frozen (DINOv3, BioClinical ModernBERT) and only small projection heads train on top, which is what lets the whole thing run on one consumer GPU. We tested retrieval in both directions on the CPJUMP1 benchmark: 51 plates, over 3 million cell images, 303 drugs, 160 genes. Plate-to-plate variation is corrected in the embedding space, so the model is not just learning which plate an image came from. Joint work with two Northeastern co-authors; I led the model design, training, and evaluation.

@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 one feature of a face and it tends to beautify the whole thing, which is a real problem when the picture is meant to be a surgical preview. The question here is how much of that you can fix from the client side, without touching the model. I compared prompt-only steering, masked compositing, and model-based inpainting across six commercial editors and one inpainting model, over 196 facelift and rhinoplasty edits, scoring how well identity survived (ArcFace) and how well the edit stayed where it was asked to stay. Plain masked compositing localized better than model-based inpainting.

@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}
}

An EfficientNet V2 classifier over all nine ISIC 2019 diagnostic categories at 384x384, trained with focal loss to cope with the heavy class imbalance, reaching 85.6% accuracy and 0.856 weighted F1 on 25,331 dermoscopic images. The more interesting part is what happens after the prediction. GradCAM++ attention is broken down along the ABCDE criteria dermatologists already use, with asymmetry, border irregularity, color variation (K-means), and diameter measured straight off the lesion mask. Because those clinical features come out as numbers, how well they line up with the model's attention can be scored, so the interpretability claim rests on a metric instead of a few good-looking 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}
}