Table 1 Overview of recent deep learning methods for cryo-EM image processing tasks.

Downstream Task	Method Name	Publication	Year	RealData	Dataset	Image
Motion Correction	NoiseFlow²⁵	ICPR	2022	No	2 sync.	No Men.
Motion Enhancement	DST-net²⁶	IEEE/ACM TCBB	2024	No	7 sync.	168
Micrograph Denoising	Topaz-denoise⁷³	Nat. Commun.	2020	Yes	18 real	3,439
Micrograph Denoising	Restore⁶³	IUCrJ	2020	Yes	4 real	No Men.
Micrograph Denoising	NT2C⁶⁵	BMC Bioinform.	2022	Yes	3real+4sync.	1,472+1,750
Micrograph Filtering	Miffi²⁸	J. Struct. Biol.	2024	Yes	9 real	45,768
Micrograph Filtering	MicrographCleaner²⁷	J. Struct. Biol.	2020	Yes	16 real	539
Particle Picking	DeepPicker²⁹	J. Struct. Biol.	2016	Yes	5 real	1,804
Particle Picking	FastParticlePicker³⁰	ICAMCS	2017	Yes	3 real	300
Particle Picking	APPLEPicker³¹	J. Struct. Biol.	2018	Yes	4 real	324+
Particle Picking	crYOLO³²	Commun. Biol.	2019	Yes	45 real	840
Particle Picking	Topaz³³	Nat. Methods	2019	Yes	6 real	2,296
Particle Picking	PIXER³⁴	BMC Bioinform.	2019	Yes	5real+1sync.	476+496
Particle Picking	PARSED³⁵	BMC Bioinform.	2020	No	No Men.	No Men.
Particle Picking	DeepCryoPicker³⁶	BMC Bioinform.	2020	Yes	3 real	350
Particle Picking	CASSPER³⁷	Commun. Biol.	2021	Yes	4 real	1,875
Particle Picking	DRPnet³⁸	BMC Bioinform.	2021	Yes	1 real	50
Particle Picking	EPicker³⁹	Nat. Commun.	2022	Yes	10 real	100
Particle Picking	SynergyNet⁴⁰	BIBM	2022	Yes	3 real	No Men.
Particle Picking	CryoTransformer⁴¹	BMC Bioinform.	2024	Yes	22 real	5,172
Particle Picking	CryoMAE⁴²	WACV	2025	Yes	5 real	1,561
Particle Picking	CryoSegNet⁴³	Brief. Bioinform.	2024	Yes	22 real	4,948

Methods are categorized and ordered hierarchically from movie-level to micrograph-level and finally particle-level, covering motion correction, micrograph denoising, micrograph filtering, and particle picking in cryo-EM. The table highlights the scale of training datasets used by these deep learning methods. While most methods are trained on real data, the limited availability of annotated cryo-EM data restricts their training to relatively small datasets, thereby constraining their generalizability. CryoCRAB includes 746 proteins, comprising 152,385 sets of raw movie frames, which can be utilized to train foundational cryo-EM models, enabling general-purpose feature extraction for downstream tasks.

Search