Table 7 Comparison of studies on machine learning for choroidal melanoma detection.

Iddir et al.²⁹

ML algorithm using ultra-

widefield fundus imaging and

B-scan ultrasonography; 223

eyes (115 choroidal nevi,

108 μm)

AUC: 0.982 (thickness) and

0.964 (subretinal fluid);

sensitivity/specificity:

0.900/0.818-1.000/0.727

Hoffmann et al.³⁰

Deep learning software using

color fundus photographs; 762

cases of nevi and melanomas

Binary accuracy: 90.9%;

AUC: 0.99; optimal

accuracy: 95.8%

Tailor et al.³⁴

ML models using multimodal

imaging; 2,870 nevi with 128

confirmed transformations

XGBoost achieved AUC:

0.864–0.931; key predictors:

tumor thickness,

diameter, shape

Shakeri et al.³⁵

Transfer learning with CNNs

on 854 fundus images;

SHAP analysis

DenseNet169 accuracy:

89%; SHAP provided

interpretability

Abrahamsson et al.³²

ML algorithms for ESI-MS

data analysis; prediction of

relative response factors

Best model: MAE of 0.19

log units and Q² of 0.84 for

CE-MS ESI+

Mann et al.³³

ML and DL for peptide

measurements prediction

in MS-based proteomics

ML improved workflow

quality; outperformed

existing biomarker assays

Zabor et al.⁵⁷

Lasso logistic regression;

123 patients (61 melanoma,

62 nevus)

AUC: 0.880 (training),

0.861 (validation); key

predictors: distance to

disc, lesion height

Yao et al.⁵⁸

DL classification of 798 ultra-

widefield retinal images; color

fusion testing

Intermediate fusion optimal;

red channel superior to

green/blue

Karamanli et al.³¹

Review of 8 studies applying

AI in choroidal lesion

assessment

U-Net: 100% sensitivity;

DenseNet121: AUC 0.9781;

ResNet50: 92.65%

accuracy

Current Study

Tear sample analysis using

protein corona in gold

nanoparticles with ESI-MS;

ML and DL with CWT-

generated images

Significant intensity differences

(p < 0.001);

Random Forest: 0.959

accuracy, 0.993 AUC;

VGG16: 0.976 accuracy,

0.997 AUC