APRIL-MedSeg: A Modular Medical Image Segmentation Toolbox Embracing Modern Paradigms

APRIL-MedSeg is an open-source, configuration-driven toolbox designed to streamline 2D medical image segmentation. It addresses the fragmentation in the field by providing a unified interface that decouples complex segmentation networks into four interchangeable modules: encoder, decoder, skip connection, and bottleneck. By using a registry-based system and YAML configuration files, the framework allows researchers to rapidly prototype, benchmark, and deploy models without writing task-specific boilerplate code.

The framework is built on a modular design philosophy that treats segmentation networks as multi-scale feature routing processes. It supports 130 distinct architectures, ranging from classic CNNs and Vision Transformers to state-of-the-art State Space Models (SSMs) like Mamba and RWKV. A key feature is its extensive support for foundation models; it integrates 39 pre-trained encoders across nine medical imaging modalities, allowing users to leverage powerful, pre-learned visual representations for downstream tasks. The system also includes a comprehensive library of 81 loss functions and 24 augmentation methods, all configurable through simple YAML inheritance.

Beyond standard supervised learning, APRIL-MedSeg incorporates five advanced training paradigms to tackle common clinical challenges:

• Data-efficient learning: Supports semi-supervised and weakly supervised methods (e.g., bounding boxes, scribbles, point clicks) to mitigate the high cost of pixel-perfect annotations.
• Generalization and compression: Includes domain adaptation techniques to handle cross-center data shifts and knowledge distillation to compress large models for efficient deployment.
• Task expansion: Features text-guided segmentation, enabling models to perform inference based on radiological language descriptions rather than just mask supervision.