This guide is still under construction

Introduction

Meddlr is a Python framework built to simplify experimentation with medical image reconstruction and analysis problems.

Meddlr is designed to facilitate the experimentation lifecycle for new ML-based methods by unifying data formats and interfaces, standardizing metric computation, and distributing state-of-the-art ML methods for these problems. Meddlr is also config-driven, which can simplify launching, monitoring, and iterating on experiments.

Features

Unified Data Interfaces

Data are represented as dictionaries with keys for inputs, outputs, and metadata. This unified representation simplifies synthesizing using different datasets during training, validation, and evaluation.

Meddlr also distributes intelligent data samplers that simplify batching data based on particular metadata. For example, if all images in the batch should be of the same dimensions, the meddlr.data.samplers.GroupSampler can be used to sample data in such a manner.

Standardized Metrics

Meddlr metrics extend :module:torchmetrics to provide a standardized interface for computing metrics across image patches (e.g. per slice) or across the entire scan. It also provides pandas and dictionary export functions to simplify metric I/O on a per-example basis.

Built-in Baselines

Meddlr offers default implementations for common baselines for MRI reconstruction and image analysis (e.g. segmentation).