Table 1.

Publicly available datasets are generally small and heterogeneously annotated. Our Ark (Fig. 1) aims to aggregate numerous datasets with heterogeneous annotations to diversify patient population, accrue knowledge from diverse experts, and meet the demand by deep learning for massive annotated training data, offering superior and robust performance (Table 2, Fig. 2 and Fig. 3) yet reducing annotation cost.

Abbrev.	Dataset	Task	Usagea	(Pre)train/val/test
1. CXPT	CheXpert [4]	classify 14 thoracic diagnoses	P|F|L|B	223414/-/234
2. NIHC	NIH ChestX-ray14 [19]	classify 14 thoracic diseases	P|F|L|B	75312/11212/25596
3. RSNA	RSNA Pneumonia [1]	classify lung opacity, abnormality	P|F|L	21295/2680/2709
4. VINC	VinDr-CXR [13]	classify 6 thoracic diagnoses	P|F|L	15000/-/3000
5. NIHS	NIH Shenzhen CXR [5]	classify tuberculosis	P|F|L	463/65/134
6. MMIC	MIMIC-II [6]	classify 14 thoracic diagnosesb	P	368879/2992/5159
7. NIHM	NIH Montgomery [5]	segment lungs	F	92/15/31
8. JSRT	JSRT [17]	segment lungs, heart, clavicles	F	173/25/49
9. VINR	VinDr-RibCXR [14]	segment 20 ribs	F	196/-/49
10. SIIM	SIIM-ACR PTX [2]	classify pneumothoraxc	L	10675/-/1372

^aThe usage of each dataset in our experiments is denoted with P for pretraining, F for fine-tuning, L for linear probing, and B for bias study.

^bThe labels of CXRs in MIMIC-II are derived from their corresponding radiology reports using NegBio [15] and CheXpert [4].

^cSIIM-ACR, originally for pneumothorax segmentation, is converted into a classification task for linear probing, as CXR-FM cannot be evaluated for segmentation using its only released API.