Revisiting 2D Foundation Models for Scalable 3D Medical Image Classification

TL;DR

This paper introduces AnyMC3D, a scalable 3D medical image classifier based on 2D foundation models, addressing key pitfalls and achieving state-of-the-art results across diverse tasks with a unified framework.

Contribution

The paper presents a lightweight, scalable 3D classifier built on 2D foundation models, supporting multi-view inputs and auxiliary supervision, and provides a comprehensive benchmark analysis.

Findings

Effective adaptation unlocks foundation model potential.

General-purpose FMs can match medical-specific FMs with proper adaptation.

2D-based methods outperform 3D architectures in 3D classification.

Abstract

3D medical image classification is essential for modern clinical workflows. Medical foundation models (FMs) have emerged as a promising approach for scaling to new tasks, yet current research suffers from three critical pitfalls: data-regime bias, suboptimal adaptation, and insufficient task coverage. In this paper, we address these pitfalls and introduce AnyMC3D, a scalable 3D classifier adapted from 2D FMs. Our method scales efficiently to new tasks by adding only lightweight plugins (about 1M parameters per task) on top of a single frozen backbone. This versatile framework also supports multi-view inputs, auxiliary pixel-level supervision, and interpretable heatmap generation. We establish a comprehensive benchmark of 12 tasks covering diverse pathologies, anatomies, and modalities, and systematically analyze state-of-the-art 3D classification techniques. Our analysis reveals key insights: (1) effective adaptation is essential to unlock FM potential, (2) general-purpose FMs can match medical-specific FMs if properly adapted, and (3) 2D-based methods surpass 3D architectures for 3D classification. For the first time, we demonstrate the feasibility of achieving state-of-the-art performance across diverse applications using a single scalable framework (including 1st place in the VLM3D challenge), eliminating the need for separate task-specific models.