Sukhrob Ilyosbekov

A text-supervised contrastive model that matches Cell Painting microscopy images of drug- and gene-perturbed cells to natural-language treatment descriptions. Frozen vision and language backbones (DINOv3, BioClinical ModernBERT) are paired with trainable projection heads, so the whole model trains on a single consumer GPU. Evaluated on bidirectional image-to-text and text-to-image retrieval over the CPJUMP1 benchmark (51 plates, 3M+ cell images, 303 drugs, 160 genes), with batch correction in the embedding space to remove plate-to-plate variation. Collaborative work with 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}
}

How much of a black-box image editor's behaviour can be controlled from outside the model? Cosmetic-surgery previews are the test case: commercial APIs beautify the whole face when asked to change a single feature. This paper compares prompt-only steering, masked compositing, and model-based inpainting across six commercial editors and one inpainting model on 196 facelift and rhinoplasty edits, scoring identity preservation with ArcFace alongside localization accuracy. Masked compositing beat model-based inpainting on localization.

@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 trained across all nine ISIC 2019 diagnostic categories at 384x384 resolution, reaching 85.6% accuracy and 0.856 weighted F1 on 25,331 dermoscopic images, with focal loss against the heavy class imbalance. The contribution is interpretability: GradCAM++ attention is decomposed along the ABCDE criteria dermatologists already use, quantifying asymmetry, border irregularity, colour variation via K-means, and diameter directly from the lesion mask. Alignment metrics between model attention and those extracted clinical features let the interpretability claim be measured rather than argued from a handful of example 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}
}