ShapeKit

TL;DR

ShapeKit is a shape-focused toolkit that significantly improves anatomical shape accuracy in medical segmentation, outperforming model architecture modifications without requiring re-training.

Contribution

We introduce ShapeKit, a practical and easy-to-integrate toolkit that enhances segmentation accuracy by focusing on shape refinement, demonstrating substantial gains over traditional model modifications.

Findings

ShapeKit improves segmentation accuracy by over 8%.

Model architecture changes yield less than 3% improvement.

ShapeKit is flexible and easy to integrate.

Abstract

In this paper, we present a practical approach to improve anatomical shape accuracy in whole-body medical segmentation. Our analysis shows that a shape-focused toolkit can enhance segmentation performance by over 8%, without the need for model re-training or fine-tuning. In comparison, modifications to model architecture typically lead to marginal gains of less than 3%. Motivated by this observation, we introduce ShapeKit, a flexible and easy-to-integrate toolkit designed to refine anatomical shapes. This work highlights the underappreciated value of shape-based tools and calls attention to their potential impact within the medical segmentation community.