Skip to content
View PhilipGraemer's full-sized avatar

Block or report PhilipGraemer

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
PhilipGraemer/README.md

Philip Graemer

PhD student in deep learning for biomedical microscopy at the University of Strathclyde and at CeMi jointly with the University of Glasgow, supervised by Dr Giuseppe Di Caprio.

I study how vision architectures learn, transfer, and fail when moved from natural images into biomedical microscopy.

My current work shows that previously reported CNN superiority over Vision Transformers on single-cell classification tasks can be an artefact of uncontrolled pretraining, and that standard optimisation recipes can unexpectedly hurt under domain shift. I also found that ViTs such as EVA-02 make excellent distillation teachers for small deployable models, seemingly imparting better dark knowledge than either CNNs or mixed councils of ViTs and CNNs.

A second strand of my research concerns microscopy and cell biology more directly, especially phase contrast imaging, quantitative phase imaging (QPI), and the morphology and differentiation of human cell types. The broader aim is to use microscopy and machine learning to overcome bottlenecks in biomedical experiments.

Current projects

  • Controlled CNN–ViT benchmark on LIVECell: Systematic comparison of eleven architectures (CNNs, ViTs, hierarchical transformers) with matched pretraining conditions. Manuscript in preparation. 📊
  • Building new single cell datasets: Building in-house datasets of phase contrast, QPI and DIC microscopy across different microscopes, different settings, different days, different facilities and different cell types. For assessing single-cell classification more realistically and testing robustness. 🔬
  • Differentiation Time Series Prediction: Collecting data and testing pipelines for predicting differentiation of pluripotent human mesenchymal stem cells into chondrocytes, adipocytes and osteoblasts. 🧫
  • PhaseDINO: Domain-adaptive self-supervised learning for phase contrast microscopy via DINO-style pretraining.🦖
  • Cross-architecture knowledge distillation: investigating when weaker teachers can outperform stronger ones across architecture families, and how ViTs and CNNs differ in uncertainty and transfer. ⚗️

Tech

PyTorch · timm · PyTorch Lightning · NVIDIA A100 / RTX Pro 6000 · Mixed precision · AdamW · WandB

Links

Pinned Loading

  1. livecell-classification-benchmark livecell-classification-benchmark Public

    Controlled benchmark comparing CNN and Vision Transformer architectures for single-cell classification on LIVECell. Multiple architectures, multi-seed reproducibility.

    Python

  2. livecell-distillation livecell-distillation Public

    Cross-architecture knowledge distillation. Temperature-scaled soft labels, combined hard/soft loss, council distillation. Didactic codebase with extensive inline documentation.

    Python

  3. segmentation-crop-checker segmentation-crop-checker Public

    Validation tool for single-cell crops extracted from segmentation masks. Flags multi-cell crops, boundary truncation, size outliers, and mask coverage issues.

    Python