Table 4.

Input data modality ablation study for skin lesion management prediction results ${\mathrm{MGMT}}_{\mathrm{pred},\mathrm{multi}}$ obtained using a multi-task model.

Experiment name	Input data			Metrics
	Clinical image	Dermoscopic image	Patient metadata	Sensitivity	Specificity	Precision	AUROC	Overall accuracy
$\mathrm{C}$	✓	✗	✗	0.5997	0.7911	0.5466	0.7781	0.6051
$\mathrm{CM}$	✓	✗	✓	0.6050	0.7983	0.5684	0.7852	0.6405
$\mathrm{D}$	✗	✓	✗	0.6935	0.8384	0.6265	0.8630	0.6962
$\mathrm{DM}$	✗	✓	✓	0.7126	0.8424	0.6622	0.8644	0.7215
$\mathrm{CD}$	✓	✓	✗	0.5830	0.8060	0.7393	0.8335	0.7342
$\mathrm{CDM}$	✓	✓	✓	0.6893	0.8299	0.7134	0.8505	0.7367

Each experiment is named so as to denote the input data modalities it uses to make the management predictions, and ‘C’, ‘D’, and ‘M’ refer to clinical image, dermoscopic image, and patient metadata, respectively.