resnet18

Binary classifier for detecting cancer cell presence on micropatterns in phase contrast microscopy images.

Model description

Fine-tuned ResNet-18 (microsoft/resnet-18) that classifies 77x77 pixel grayscale crops as present (cancer cell on micropattern) or absent (empty micropattern). Part of the pipeline for analyzing T-cell killing assays.

Training

• Base model: microsoft/resnet-18
• Dataset: 420 manually annotated samples (28 crops x 15 timepoints) from a single position (Pos150)
• Class balance: 60% present / 40% absent
• Epochs: 20
• Batch size: 32
• Learning rate: 1e-4
• Validation split: 20%
• Training time: ~45 seconds on Apple Silicon

Performance

Metric	Value
Accuracy	96.5%
F1 Score	0.97

Control validation (Pos140, no T-cells added)

• 47/53 cells correctly classified as surviving all 50 timepoints
• 6 false deaths (~11% false positive death rate)

Usage

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image

processor = AutoImageProcessor.from_pretrained("keejkrej/resnet18")
model = AutoModelForImageClassification.from_pretrained("keejkrej/resnet18")

image = Image.open("crop.png").convert("RGB")
inputs = processor(image, return_tensors="pt")
outputs = model(**inputs)

predicted_class = outputs.logits.argmax(-1).item()
label = model.config.id2label[predicted_class]  # "present" or "absent"

Input format

• 77x77 pixel grayscale microscopy crops (uint16 normalized to uint8)
• Converted to RGB (3-channel) before processing
• Single micropattern site per crop

Labels

ID	Label
0	absent
1	present