ChampKit

ChampKit facilitates systematic evaluation of deep neural networks for patch-based histopathology classification and benchmarking.

Key Features:

• Extensible Evaluation Framework: Supports comprehensive and reproducible evaluation of neural network models on histopathology datasets.
• timm-supported Architectures: Leverages architectures available through timm for model selection and evaluation.
• API Support for External Models: Provides an API to integrate external models for evaluation alongside built-in architectures.
• Dataset Curation: Curates a broad range of public histopathology datasets for benchmarking.
• Baseline Performance Establishment: Establishes baseline metrics for ResNet18, ResNet50, and R26-ViT across multiple datasets.
• Transfer Learning Analysis: Systematically compares models trained from random initialization with those using ImageNet and self-supervised pretrained weights.

Scientific Applications:

• Model selection for histopathology: Identifying optimal neural network architectures for patch-based histopathology classification tasks.
• Benchmarking across datasets: Comparative evaluation of deep learning models across multiple public histopathology datasets.
• Transfer learning assessment: Quantifying the impact of ImageNet and self-supervised pretraining on histopathology model performance.

Methodology:

Training and evaluation of deep learning models (including ResNet18, ResNet50, R26-ViT and timm-supported architectures) on curated histopathology datasets, comparing random weight initialization versus ImageNet and self-supervised pretrained weights, and reporting baseline performance metrics.