CellVision

Microscopy Cell Type Classifier

A PyTorch pipeline that classifies seven peripheral-blood cell types from microscopy images, with stain normalization and GradCAM interpretability behind a Streamlit interface.

Methods & contributions

• Seven cell types — RBC, neutrophil, lymphocyte, monocyte, eosinophil, basophil, and platelet, with morphology-grounded descriptions.
• Two architectures — a custom CellNet CNN (4 conv blocks, global average pooling) and ResNet18 transfer learning.
• Interpretability — GradCAM activation maps over the final convolutional layer.
• Preprocessing — stain normalization and Otsu segmentation; logits feed CrossEntropyLoss with softmax at inference only.

Data flow

smear image → stain normalization → CNN / ResNet18
           → softmax over 7 classes → GradCAM overlay

Technologies