2.5 CNN: Leveraging 2D CNNs to Pretrain 3D Models in Low-Data Regimes for COVID-19 Diagnosis

Abstract

Accurately diagnosing COVID-19 from three-dimensional (3D) Computed Tomography (CT) scans can be challenging due to the high dimensionality of volumetric data and the scarcity of annotated samples in many clinical datasets. We propose a two-stage (“2.5D”) approach that first trains a 2D convolutional neural network (CNN) on individual CT slices, thereby expanding the training set and mitigating data limitations. We then reuse the feature extraction layers of this 2D model in a second stage by stacking slice-level embeddings and training a lightweight 3D classifier on top. This design combines the benefits of slice-level representation learning with the volumetric context essential for medical image interpretation. Evaluations on the MosMed dataset (1130 CT scans) show that our pipeline achieves a weighted accuracy of 94.73% and an unweighted accuracy of 95.35%, surpassing purely 2D and purely 3D methods. Additionally, we examine tasks that differentiate between various COVID-19 severity levels, demonstrating robust performance under notable class imbalance. Finally, we outline theoretical and algorithmic considerations, including how the 2.5D approach relates to multi-instance learning frameworks and how it can reduce complexity relative to naive 3D training in low-data regimes.