TABLE 4.

Overview of RL in medical image classification

Author	ROI	Modality	Algorithm	State	Action number	Reward design
60	Skin	Dermascope	DQN	Patient's history answers + output probability of pretrained CNN	300	Probability of correct condition if asked the question
58	Chest	Xray	Policy Gradient	*	*	*
30	Brain	MR	DQN + TD	Image overlaid in red or green	2	Classification correctness
55	Brain	MR	DQN + TD	Image overlaid in red or green	2	Classification correctness
57	Breast	Ultrasound	REINFORCE	Weights	3	Classification correctness
57	Chest	CT	DDPG	Predictions of the unlabeled images	Continuous	Possibility of being classified incorrectly
52	Cervix, Lymph Node	Histopathology	PPO	Selected images	2	Maximum validation accuracy of last epochs
70	Breast	Histopathology	Policy Gradient	Learning status representation + Incoming data statistics	2	Performance of the selection mechanism

*Indicate that the missing part is not clearly defined in the original paper.

A summary of the works we reviewed in this section is given in Table 5.