Rigorous research.
Real-world impact.

Behind Melanoma Score is a research programme led by Professor Gustav Stålhammar at St. Erik Eye Hospital and Karolinska Institutet. Here is how the model was built and validated, and what it means for patients, healthcare providers, and optometrists.

See the research

Find the lesion
Score from 0 to 1

Figure 1. Simplified from the method in Ophthalmology Science, 2024, where 0.63 was the operating point, as in TVST, 2025. The lesion is painted, not a patient's, so no score is shown.

Figure 1, an illustration. An image of the back of the eye, with a painted pigmented lesion. A circle centres on the lesion, and the model's method, as described in Ophthalmology Science, 2024 and simplified here to two steps, plays out. First, an outline is drawn around the lesion and the circle becomes a square centred on it. Second, the outline unrolls into a scale from 0 to 1 under the square, with a mark at 0.63, the operating point used in the Ophthalmology Science 2024 and TVST 2025 studies. No score is shown, because the lesion is painted and not a patient's.

.re-track { height: auto !important; } .re-stage { position: relative !important; height: auto !important; min-height: 92vh; } .re-field { display: none !important; } .re-under { opacity: 1 !important; }

At a glance

The short version, in nine lines.

Each line links to its section. Every figure names its source.

Developed by: Prof. Gustav Stålhammar's research group at St. Erik Eye Hospital and Karolinska Institutet, with Mats Holmström. Ten authors across three papers.
Intended use: Decision support for clinicians who assess pigmented lesions at the back of the eye, from opticians and optometrists to ocular oncologists. Cancers, 2026
Input: One colour photograph of the back of the eye. The studies used standard field and ultra-widefield photographs from two camera systems. Ophthalmology Science, 2024
Output: A risk score from 0 to 1. The same image gives the same score every time, whoever runs it. Cancers, 2026
Data: In the 2024 study, 802 photographs from 688 patients, each assessed by an ocular oncologist. Harmless meant at least five years without being rediagnosed as melanoma. Ophthalmology Science, 2024
Tested: Three peer-reviewed studies: against twelve ophthalmologists (Ophthalmology Science, 2024), 29 opticians and optometrists on 25 photographs with five melanomas (TVST, 2025), and the MOLES checklist in expert hands (Cancers, 2026).
Key result: In the 2024 study, AUC 0.88 on both test sets. At the 0.63 threshold, 100 % sensitivity with 74 % specificity (internal set) and 80 % with 81 % (external set). Ophthalmology Science, 2024
Still open: Tested on images from several hospitals and validated on an external test set. Prospective multicentre studies are planned.
Disclosures: Each paper names its funding. The 2026 paper declares that Mats Holmström and Gustav Stålhammar are founders of Eyedentity AB and hold equity in the company. Cancers, 2026

01Studies

Developed, compared, and tested where referrals begin.

From development and validation, to optician triage, to a head-to-head comparison with the MOLES checklist.

Study 1, Ophthalmology Science, 2024: Can the model tell a small melanoma from a harmless lesion?

A retrospective multicentre study of 802 photographs from 688 patients, all examined at St. Erik Eye Hospital. The model was built on one part of them and tested on two sets it had never seen: 86 photographs held back from the same collection, and 53 taken at other clinics across Sweden before the patients came to St. Erik. Twelve ophthalmologists, from residents to ocular oncologists, read the same 86 photographs without knowing the answers.

The model, from the photograph alone
Checklists, also scored with ultrasound or OCT

Internal test set86 photographs, held back from the same collection

Model 0.88 (0.82 to 0.95)

MOLES 0.77 (0.66 to 0.88)

TFSOM-UHHD 0.67 (0.54 to 0.81)

Both checklists were scored by ophthalmologists who also had ultrasound or OCT from the records. The model had the photograph only.

External test set53 photographs, other clinics in Sweden

Model 0.88 (0.74 to 1.00)

The ophthalmologists read the internal set only, and the checklists were compared with the model on the internal set only.

0.50.60.70.80.91.0

AUC

AUC, how often a score ranks a melanoma above a harmless lesion: 0.5 is a coin toss, 1.0 is perfect. Whiskers show 95 % confidence intervals. Values from the paper's Figure 4 and its results.Ophthalmology Science, 2024

Figure 2. AUC on the two test sets
Test set	Method	AUC	95 % confidence interval
Internal test set, 86 photographs, held back from the same collection	The model, from the photograph alone	0.88	0.82 to 0.95
Internal test set, 86 photographs, held back from the same collection	MOLES, Checklists, also scored with ultrasound or OCT	0.77	0.66 to 0.88
Internal test set, 86 photographs, held back from the same collection	TFSOM-UHHD, Checklists, also scored with ultrasound or OCT	0.67	0.54 to 0.81
Internal test set, 86 photographs, held back from the same collection	Both checklists were scored by ophthalmologists who also had ultrasound or OCT from the records. The model had the photograph only.
External test set, 53 photographs, other clinics in Sweden	The model, from the photograph alone	0.88	0.74 to 1.00
External test set, 53 photographs, other clinics in Sweden	The ophthalmologists read the internal set only, and the checklists were compared with the model on the internal set only.

At the 0.63 threshold, set before testing, the model had 100 % sensitivity and 74 % specificity on the internal set, and 80 % sensitivity and 81 % specificity on the external set. On the internal set, the one the twelve ophthalmologists also read, its sensitivity was significantly higher than the residents' and the consultants' and not significantly different from the ocular oncologists'; its specificity did not differ significantly from any group's. Ophthalmology Science, 2024

The authors' conclusion

“This study develops and validates a deep learning algorithm for differentiating small choroidal melanomas from nevi, matching or surpassing the discriminatory performance of experienced human ophthalmologists.”

Ophthalmology Science, 2024

The authors' caveat

“Further research will aim to validate its utility in clinical settings.”

Ophthalmology Science, 2024

Sensitivity: The share of melanomas correctly classed as melanoma.
Specificity: The share of harmless lesions correctly classed as harmless.

Internal test set, 86 photographs, with the model at its 0.63 threshold

Reader	Sensitivity	Specificity	Dunn P against the model, sensitivity and specificity
Model	100 %	74 %
Resident ophthalmologists, 6	85 %	63 %	0.04 and 0.08
Consultant ophthalmologists, 3	83 %	78 %	0.006 and 0.36
Ocular oncologists, 3	98 %	70 %	> 0.99 and > 0.99

Dunn post hoc tests with Bonferroni correction. Across the four groups, sensitivity differed (Kruskal-Wallis P < 0.001) and specificity did not (P = 0.26). All twelve ophthalmologists pooled against the model (Mann-Whitney U): sensitivity P = 0.006, specificity P = 0.54.

Ophthalmology Science, 2024

AUC on the internal test set, 86 photographs

Method	Scored from	AUC (95 % CI)	Against the model
Model	The photograph	0.88 (0.82 to 0.95)
MOLES	The photograph plus ultrasound and/or OCT from the records	0.77 (0.66 to 0.88)	DeLong P < 0.001
TFSOM-UHHD	The photograph plus ultrasound and/or OCT from the records	0.67 (0.54 to 0.81)	DeLong P < 0.001

Ophthalmology Science, 2024

By camera type, internal test set

Camera	Photographs	AUC (95 % CI)	Sensitivity	Specificity
Ultra-widefield, Optos, 200°	31	0.89 (0.76 to 1.00)	100 %	62 %
Standard field, Canon, 45°	55	0.89 (0.81 to 0.97)	100 %	77 %

The external set gave AUC 0.88 (0.74 to 1.00), with 80 % sensitivity and 81 % specificity. The ophthalmologists did not read it.

Ophthalmology Science, 2024

Study 2, TVST, 2025: How do opticians triage small pigmented lesions, and how does the model's threshold compare on the same images?

Twenty-nine Swedish opticians and optometrists graded 25 photographs of small pigmented lesions, five of them melanomas. They scored each one with MOLES, the checklist the Swedish referral guideline uses, and decided whether to refer. The study then compares their referrals with the model's threshold on the same images.

Opticians, the referrals they recorded
The model's threshold, calculated on the same images

Harmless lesions20 images

Model 10 %

Opticians 68 %

The opticians' 580 decisions included 397 referrals; the model's threshold flagged 2 of the 20 images.

Melanomas5 images

Model 80 %

Opticians 95 %

The opticians' 145 decisions included 138 referrals; the model's threshold flagged 4 of the 5 images. It missed one of the five.

0 %25 %50 %75 %100 %

Share referred

Opticians: the referrals they recorded, 29 decisions per image. Model: its own classification of each image at the 0.63 threshold, calculated on the same images. The opticians' own decisions after seeing the score were not recorded.TVST, 2025

Figure 3. Share referred, same 25 images
Lesions	Opticians or model	Share referred
Harmless lesions, 20 images	The model's threshold, calculated on the same images	10 %
Harmless lesions, 20 images	Opticians, the referrals they recorded	68 %
Harmless lesions, 20 images	The opticians' 580 decisions included 397 referrals; the model's threshold flagged 2 of the 20 images.
Melanomas, 5 images	The model's threshold, calculated on the same images	80 %
Melanomas, 5 images	Opticians, the referrals they recorded	95 %
Melanomas, 5 images	The opticians' 145 decisions included 138 referrals; the model's threshold flagged 4 of the 5 images. It missed one of the five.

In the authors' words, the opticians' low specificity “reflects guideline structure rather than clinical incompetence”: the guideline asks them to refer at a MOLES score of 1 or more, and a score of 1 is what MOLES itself calls a low-risk nevus. The authors write that the model's threshold on the same images would have “reduced false-positive referrals 10-fold”. With five melanomas in the set, that came at the cost of one: the threshold flagged 4 of 5, where the opticians' recorded decisions referred 95 %. TVST, 2025

From the opticians' questionnaire

In the study's first step, 54 opticians and optometrists from 17 shops answered a questionnaire; 29 graded the images in the second.

83 %find it hard to judge whether the referral criteria are metn = 54
65 %typically take two minutes or more to assess a lesionn = 54
70 %see five or more customers with pigmented lesions a yearn = 54
85 %have a fundus camera where they workn = 54

TVST, 2025

The authors' conclusion

“Swedish opticians and optometrists detect small pigmented choroidal lesions with high sensitivity but poor specificity, leading to frequent, unnecessary referrals.”

TVST, 2025

The authors' caveat

“Notably, among these 25 lesions, only 5 were melanomas, meaning that a single missed melanoma reduced sensitivity from 100 % to 80 %.”

TVST, 2025

The guideline's rule and a stricter cut-off, applied to the opticians' MOLES scores, against the model's threshold. Same 25 images, 5 melanomas and 20 harmless lesions

Rule	Sensitivity	Specificity	Accuracy	PPV	NPV
MOLES 1 or more, the guideline	98 % (SD 6)	17 % (SD 16)	33 %	23 % (SD 4)	99 % (SD 4)
MOLES 3 or more	75 % (SD 29)	53 % (SD 28)	57 %	34 % (SD 16)	93 % (SD 7)
Model above 0.63	80 %	90 %	88 %	67 %	95 %

Opticians: mean and standard deviation across the 29. The model has no spread, because “the algorithm produces an identical score with each image analysis”. At the 0.63 threshold the model flagged 4 of the 5 melanomas and 2 of the 20 harmless lesions. The paper gives no confidence intervals for these figures. PPV is the share of referred lesions that were melanomas, NPV the share of lesions not referred that were harmless; both depend on the share of melanomas, here 5 of 25.

TVST, 2025

The 98 % and 17 % describe the guideline's rule applied to the opticians' MOLES scores. The decisions they recorded sent on 95 % of melanomas and 68 % of harmless lesions, as in Figure 3. TVST, 2025

“Quadrupled the odds of correctly referring a melanoma” compares two odds ratios: about 9.0 for the opticians' recorded referrals and 36.6 for the model's classification. It measures how well referrals separate melanomas from harmless lesions, not how many melanomas were found. “10-fold” in the abstract and “to one-tenth” in the conclusions are the authors' summary of how far false-positive referrals would have fallen had the referrals followed the model's threshold. TVST, 2025

Net benefit at each referral threshold (decision curve analysis)

Threshold	Opticians	Model
10 %	0.1	0.2
20 %	0.0	0.1
30 %	−0.1	0.1
40 %	−0.3	0.1
50 %	−0.4	0.1
60 %	−0.6	0.0
70 %	−0.8	0.0
80 %	−1.2	−0.2
90 %	−1.6	−0.6

The threshold is the lowest probability of melanoma at which a lesion would be referred. The model's net benefit was higher than the opticians' at every threshold; theirs turned negative from 30 %, the model's from 80 %.

TVST, 2025

Each optician's MOLES scores gave an AUC between 0.61 and 0.88, mean 0.76. The model's was 0.80, numerically higher than 22 of the 29. After correction for 29 comparisons, no single difference was statistically significant. TVST, 2025

Study 3, Cancers, 2026: How does the model's score compare with the MOLES checklist in expert hands?

Eighty-six lesions from St. Erik Eye Hospital, 29 of them melanomas: the 2024 study's internal test set, none of it used to build the model. Ocular oncologists scored MOLES with OCT and autofluorescence at hand, the checklist's best case. The model, which the paper calls “MelAInoma”, had the colour photograph only. No threshold was applied; the score was used on its full scale from 0 to 1.

The model, from the photograph alone
MOLES, scored by ocular oncologists with OCT and autofluorescence

Standard field, 45°55 photographs

Model 0.87 (0.78 to 0.96)

MOLES 0.80 (0.67 to 0.93)

P = 0.42 for the difference (DeLong test)

Ultra-widefield, 200°31 photographs

Model 0.87 (0.72 to 1.00)

MOLES 0.79 (0.61 to 0.97)

P = 0.48 for the difference (DeLong test)

0.50.60.70.80.91.0

AUC

The differences between the two were “not statistically detectable”, in the authors' words. Whiskers show 95 % confidence intervals.Cancers, 2026

Figure 4. AUC by camera type
Camera	Method	AUC	95 % confidence interval
Standard field, 45°, 55 photographs	The model, from the photograph alone	0.87	0.78 to 0.96
Standard field, 45°, 55 photographs	MOLES, scored by ocular oncologists with OCT and autofluorescence	0.80	0.67 to 0.93
Standard field, 45°, 55 photographs	P = 0.42 for the difference (DeLong test)
Ultra-widefield, 200°, 31 photographs	The model, from the photograph alone	0.87	0.72 to 1.00
Ultra-widefield, 200°, 31 photographs	MOLES, scored by ocular oncologists with OCT and autofluorescence	0.79	0.61 to 0.97
Ultra-widefield, 200°, 31 photographs	P = 0.48 for the difference (DeLong test)

The model's score explained more of the variation in diagnosis than MOLES did, Tjur R² 0.38 against 0.27. The two scores share only 16 % of their variance, so they often rate the same lesion differently. Cancers, 2026

The authors' conclusion

“MelAInoma demonstrates slightly stronger association with melanoma diagnosis and provides fully reproducible output, supporting its role as a complementary aid in lesion triage.”

Cancers, 2026

The authors' caveat

“Statistical association does not by itself establish clinical utility, which depends on referral thresholds, decision context, and acceptable trade-offs between sensitivity and specificity.”

Cancers, 2026

Conflict of interest, as declared in the paper

“Mats Holmström and Gustav Stålhammar are founders of Eyedentity AB and hold equity in the company, which develops the MelAInoma software evaluated in this study. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.”

Cancers, 2026

AUC: The chance that a score ranks a randomly chosen melanoma above a randomly chosen harmless lesion. 0.5 is a coin toss, 1.0 is perfect.
Tjur R²: How well a score separates melanomas from harmless lesions when used alone to predict them: the gap between the two groups' average predicted chance of melanoma. 0 means no separation, 1 complete separation.

Each score alone, outcome melanoma, 86 lesions

Score	Step	Odds ratio (95 % CI)	Tjur R²
Model score	per 0.1	2.27 (1.57 to 4.23)	0.38
MOLES	per point	2.29 (1.59 to 3.50)	0.27

Both P < 0.0001. The steps differ, 0.1 on the model's 0 to 1 scale against one point on the MOLES scale of 0 to 10, so the two odds ratios are not compared directly.

Cancers, 2026

Both scores in one regression

Score	Step	Odds ratio (95 % CI)	P
Model score	per 0.1	2.82 (1.62 to 6.85)	< 0.0001
MOLES	per point	2.24 (1.40 to 4.02)	0.002

Each score stayed associated with melanoma when the other was taken into account. After bootstrap correction this regression had an AUC of 0.92 (0.93 before correction) and a calibration slope of 0.89, which the authors read as “modest overfitting but overall stable model performance”.

Cancers, 2026

The two scores rose together, but MOLES explained only 16 % of the variation in the model's score (R² 0.16). The authors conclude that the two “will often assign different risk levels to the same lesion and therefore should not be considered interchangeable”. Cancers, 2026

This paper applies no threshold and reports no sensitivity or specificity of its own. Cancers, 2026

The sensitivity and specificity figures it cites, among them 80 % and 90 % for the model at the 0.63 threshold, are restated from the 2025 opticians' study in TVST, where the threshold was applied to 25 photographs, only five of them melanomas. TVST, 2025

On these same 86 photographs, the 2024 study reported an AUC of 0.89 for each camera type, with the model it describes. This paper does not compare the two sets of figures. Ophthalmology Science, 2024

02Method

The model finds the lesion first, then scores it.

Here is how the model in Figure 1 gets from photograph to answer, as the 2024 paper describes it.

Find the lesion
The model marks the pixels in the photograph that belong to the lesion.
Crop around it
It centres a square crop on the lesion, so the next step reads the lesion rather than the whole retina.
Score from 0 to 1
It reads the crop and returns one risk score between 0 and 1. A threshold turns the score into a recommendation: refer, or do not refer. The 2024 study chose one threshold, 0.63, on the training and tuning images and kept it unchanged for both test sets. The 2025 study used the same threshold.

In the 2026 paper's words, the model “provides fully reproducible output, yielding the same score each time a given image is analyzed, independent of the user”

Cancers, 2026

Every label came from a specialist's examination.

The 2024 study drew on photographs from multiple institutions across Sweden, collected at St. Erik Eye Hospital's ocular oncology service, the national referral centre where every patient was examined.

Figure 5. How the 2024 study's 802 photographs were split

Training495 images, 410 patients
Tuning168 images, 139 patients
Internal test86 images, 86 patients
External test53 images, 53 patients, other clinics

No image was used in more than one set, and no patient crossed between sets. The threshold was chosen on the training and tuning sets (the paper calls the tuning set the validation set); the two test sets were only used to test.Ophthalmology Science, 2024

How each image got its answer

Every patient was examined by an ocular oncologist, with photographs, OCT, ultrasound and slit lamp, and a biopsy in the rare case where the examinations were inconclusive.
A lesion counted as harmless only after at least five years without being rediagnosed as melanoma.
A lesion counted as melanoma if it was diagnosed as one, then or later, for example after it grew. The authors added this rule to help the model pick up early signs of malignancy, at a time when a small melanoma is hard to tell from a harmless lesion.

Ophthalmology Science, 2024

03Intended use

One question, asked of one photograph.

Is a pigmented lesion at the back of the eye a harmless nevus, or a small melanoma? The model gives a risk score for that one question, from a colour photograph.

Such lesions are common. Up to 700,000 people in Sweden have one, according to SVT. Almost all stay harmless, but a small melanoma can look just like one.

Telling them apart is hard even for experienced ocular oncologists, and a growing number are found by chance: in diabetic eye screening, before cataract surgery, and in the fundus photographs opticians and optometrists take, as the opticians' study in TVST (2025) describes.

4 % to 7 % of Caucasian adults have a choroidal nevus, according to the population studies the 2024 paper cites. Among people of African and Asian descent it is under 1 %.

Ophthalmology Science, 2024

“The intended use is decision support for clinicians who assess pigmented choroidal lesions with varying frequency and expertise, including opticians, optometrists, general ophthalmologists, and subspecialized ocular oncologists.”

Cancers, 2026

A risk assessment that supports a referral decision. The clinician decides.

04Limits

What the studies leave open.

Every paper names its own limits. Here they are side by side, each with the authors' own words.

Referred patients

Every patient was examined at one national referral centre, so the mix of cases is a specialist's, not an optician's. Only the 2024 study had a test set from other clinics; the 2026 paper names its single centre as a limit.

“conducted at a single national ocular oncology referral center, which may limit generalizability” Cancers, 2026

Two cameras

Only two kinds of camera were used, one standard field and one ultra-widefield.

“we advise against its application without prior adaptation and revalidation on other types of images” Ophthalmology Science, 2024

Small sets

The test sets are small, and the opticians' study held five melanomas.

“a single missed melanoma reduced sensitivity from 100 % to 80 %” TVST, 2025

Clinical answers

The reference answers came from ocular oncologists' clinical assessment, with a biopsy only in rare unclear cases. In the 2026 study, no lesion was confirmed by tissue.

“none of the included lesions underwent histopathologic confirmation” Cancers, 2026

Past photographs

All three studies used photographs taken in the past, and none recorded a decision made with the score in hand.

“Further investigation is needed, ideally in prospective cohorts, to validate the practical application of this algorithm in clinical settings” Ophthalmology Science, 2024

Ophthalmology Science, 2024

The model can only be as right as its reference: “our algorithm's accuracy is contingent upon the precision of the ocular oncologists' clinical evaluations”.
Three ocular oncologists took part as readers, which “may not necessarily reflect the broader ocular oncology community”.
The oncologists' lean toward higher sensitivity and lower specificity “could stem from their routine access to a broad array of diagnostic tools”.
An end-to-end model was tried in planning and set aside: “it was not feasible to train a model in this manner, given the amount of data we have”.
The sample size “utilized for the development, validation, and testing of the algorithm was limited”, and with larger samples “the sensitivity and specificity metrics might have varied”.
Lesions that look unusual may be read less well: “its efficacy might diminish for lesions with atypical appearances not represented in the training dataset”.
Training used two imaging systems only, “leaving its effectiveness on alternative platforms untested”, and ultra-widefield photographs “may not reproduce true colors accurately”.
Leaving out poor photographs “could have introduced a bias favoring images from centers more experienced in examining fundus lesions”, though the authors “did not observe lower image quality in the external test cohort”.
Lesions hidden by bleeding or detachment were left out, so the model was not trained on them, “which may lead to misclassification”.

Ophthalmology Science, 2024

TVST, 2025

The set held “only 25 pigmented fundus lesion images”, five of them melanomas.
Answers were anonymised, so the authors could not “correlate individual diagnostic performances” with the questionnaire.
The threshold was kept from the 2024 study, set with equal weight on sensitivity and specificity, although “maximizing sensitivity is paramount in the screening setting faced by opticians and optometrists”.
The reference answers were clinical, and “exceptions do occur, particularly with small lesions”.
No genetic or other prognostic data were available, because every melanoma was treated promptly.
Not listed by the authors, but part of the design: the same 25 images were used in both phases, and the opticians' decisions with the score were not recorded, so the comparison with the model is a calculation.

TVST, 2025

Cancers, 2026

The study was “conducted at a single national ocular oncology referral center, which may limit generalizability”.
No lesion was confirmed by tissue: “none of the included lesions underwent histopathologic confirmation”.
Two camera systems “may have introduced variability in image properties”.
Molecular markers could not be studied: “genetic and cytogenetic data were not available”.
The reference answers and the MOLES scores came from the same specialist service, a “risk of shared-feature incorporation” that the authors expect would, if anything, “favor MOLES over MelAInoma”.
The calibration test “has limited power in small samples”.
The MOLES categories were imbalanced, which “may reduce statistical stability and widen confidence intervals”.
MOLES had OCT and autofluorescence, while the model “was restricted to color fundus photography”.
“Non-expert performance was not evaluated in this cohort.”

Cancers, 2026

05People

The people behind the evidence.

Prof. Gustav Stålhammar

Professor of ophthalmology at Karolinska Institutet
Senior consultant and ocular oncologist at St. Erik Eye Hospital

He leads the research behind the model and is the corresponding author on all three papers. At St. Erik Eye Hospital he is one of a handful of specialists who treat the whole country's cases of uveal melanoma. He is also a founder of Eyedentity and its Chief Scientific Officer.

100+: peer-reviewed articles by Prof. Stålhammar, three of them on the melanoma model
#4: in the world in uveal melanoma research over the past five years, ranked by ScholarGPSSee the ranking (opens in a new tab)

Prof. Gustav Stålhammar in blue scrubs, arms crossed

Authors on the three papers

Jakub BartoszekKalmar County HospitalOphthalmology Science, 2024
Hannah Coudé AdamKarolinska Institutet and St. Erik Eye HospitalCancers, 2026
Maria FiliSt. Erik Eye Hospital and Karolinska InstitutetOphthalmology Science, 2024
Mats HolmströmEyedentity AB, as printed on the 2026 paperOphthalmology Science, 2024 TVST, 2025 Cancers, 2026
Marco A. Lumia MichalskiSt. Erik Eye Hospital and Karolinska InstitutetOphthalmology Science, 2024 TVST, 2025
Maria NilssonKarolinska InstitutetTVST, 2025
Shiva SabazadeSt. Erik Eye Hospital and Karolinska InstitutetOphthalmology Science, 2024 TVST, 2025
Gustav StålhammarSt. Erik Eye Hospital and Karolinska InstitutetOphthalmology Science, 2024 TVST, 2025 Cancers, 2026
Katerina StriplingKarolinska Institutet and St. Erik Eye HospitalCancers, 2026
Sara WittskogSynologen ABTVST, 2025

Most of the authors are based at St. Erik Eye Hospital, in its Ocular Oncology Service and Ophthalmic Pathology Laboratory, and at Karolinska Institutet's Division of Eye and Vision, with co-authors from Kalmar County Hospital on the 2024 paper, Synologen AB on the 2025 paper and Eyedentity AB on the 2026 paper. The first studies on the model were published before the company existed.

Fili M, Ermedahl Conradi M, Klaff Dahl M, Sabazade S, Casselholm De Salles M, Stålhammar G. “Long-Term Local Control and Mortality After Transpupillary Thermotherapy of Small Uveal Melanomas.” Clinical & Experimental Ophthalmology, 2026. PMID 42104514. doi.org/10.1111/ceo.70126, opens in a new tab
Nissen K, Kiilgaard JF, Fili M, Seregard S, Navaratnam J, Krohn J, Pedersen Bærland T, Eid Robsahm T, Eide N, Stålhammar G. “Increasing Incidence of Posterior Uveal Melanoma in Scandinavia 1960-2022: A Tri-National Study.” American Journal of Ophthalmology, 2025. PMID 40054546. doi.org/10.1016/j.ajo.2025.03.002, opens in a new tab
Stålhammar G, Hagström A, Ermedahl Conradi M, Williams PA. “Choroidal nevi and melanoma doubling times and implications for delays in treatment: A systematic review and meta-analysis.” Survey of Ophthalmology, 2025. PMID 39343315. doi.org/10.1016/j.survophthal.2024.09.004, opens in a new tab
Stålhammar G, Coupland SE, Ewens KG, Ganguly A, Heimann H, Shields CL, Damato B. “Improved Staging of Ciliary Body and Choroidal Melanomas Based on Estimation of Tumor Volume and Competing Risk Analyses.” Ophthalmology, 2024. PMID 38071620. doi.org/10.1016/j.ophtha.2023.10.026, opens in a new tab
Stålhammar G. “Delays between Uveal Melanoma Diagnosis and Treatment Increase the Risk of Metastatic Death.” Ophthalmology, 2024. PMID 37995852. doi.org/10.1016/j.ophtha.2023.11.021, opens in a new tab
Stålhammar G, Kalirai H, Coupland SE, Grisanti S, Sonntag SR, Tura A, Eleuteri A, Hussain RN, Heimann H, Coudé Adam H. “Metastatic potential of small posterior uveal melanomas.” Acta Ophthalmologica, 2026. PMID 40838555. doi.org/10.1111/aos.17582, opens in a new tab
Asplund E, Fili M, Pansell T, Brautaset R, Nilsson M, Stålhammar G. “The prognostic implication of visual acuity at the time of uveal melanoma diagnosis.” Eye, 2023. PMID 36434284. doi.org/10.1038/s41433-022-02316-8, opens in a new tab
Gill V, Herrspiegel C, Sabazade S, Fili M, Bergman L, Damato B, Seregard S, Stålhammar G. “Trends in Uveal Melanoma Presentation and Survival During Five Decades: A Nationwide Survey of 3898 Swedish Patients.” Frontiers in Medicine, 2022. PMID 35721086. doi.org/10.3389/fmed.2022.926034, opens in a new tab
Sabazade S, Herrspiegel C, Gill V, Stålhammar G. “No differences in the long-term prognosis of iris and choroidal melanomas when adjusting for tumor thickness and diameter.” BMC Cancer, 2021. PMID 34819035. doi.org/10.1186/s12885-021-09002-0, opens in a new tab
Sun M, Zhou W, Qi X, Zhang G, Girnita L, Seregard S, Grossniklaus HE, Yao Z, Zhou X, Stålhammar G. “Prediction of BAP1 Expression in Uveal Melanoma Using Densely-Connected Deep Classification Networks.” Cancers, 2019. PMID 31623293. doi.org/10.3390/cancers11101579, opens in a new tab

06Papers

Read the papers in full.

All three are open access under CC BY 4.0. The TVST and Cancers titles use the model's research name; the Ophthalmology Science paper does not name it.

August 2024, Ophthalmology Science, 2024

June 2025, TVST, 2025

March 2026, Cancers, 2026

August 2024Open access
Development and Validation of a Deep Learning Algorithm for Differentiation of Choroidal Nevi from Small Melanoma in Fundus Photographs
Sabazade, Lumia Michalski, Bartoszek, Fili, Holmström, Stålhammar
Ophthalmology Science, volume 5, issue 1, article 100613. Online 30 August 2024, in the 2025 volume.
doi.org/10.1016/j.xops.2024.100613, opens in a new tab Read the paper, Ophthalmology Science, 2024, opens in a new tab
PubMed
39421390
PubMed Central
PMC11483474
Funding
“Support for this study was provided to G.S. from the Swedish Society of Medicine (SLS-971390), Region Stockholm (FoUI-981345), the Swedish Cancer Society (23 2613 Fk), and the Swedish Eye Foundation (2023-05-04). The sponsors or funding organizations had no role in the design or conduct of this study.”
Ethics
Swedish Ethical Review Authority, reference 2022-06210-02. Informed consent was waived because the study was retrospective.
Conflicts
“All authors have completed and submitted the ICMJE disclosures form.” The paper's disclosure section lists the grant support above and nothing else.
June 2025Open access
Challenges for Opticians in Evaluating Small Pigmented Choroidal Lesions: Potential Support From the MelAInoma Deep Learning Algorithm
Sabazade, Lumia Michalski, Wittskog, Holmström, Nilsson, Stålhammar
Translational Vision Science & Technology, volume 14, issue 6, article 29. Online 23 June 2025.
doi.org/10.1167/tvst.14.6.29, opens in a new tab Read the paper, TVST, 2025, opens in a new tab
PubMed
40548640
PubMed Central
PMC12186828
Funding
“Supported by the Swedish Eye Foundation (Ögonfonden, 23-05-04; GS). The sponsor or funding organization had no role in the design or conduct of this study.”
Ethics
Swedish Ethical Review Authority, reference 2024-02620-02. Informed consent was waived because the study was retrospective.
Conflicts
“Disclosure: S. Sabazade, None; M.A. Lumia Michalski, None; S. Wittskog, None; M. Holmström, None; M. Nilsson, None; G. Stålhammar, None”
March 2026Open access
Comparison of MOLES and MelAInoma for Differentiating Small Choroidal Melanomas from Nevi
Stripling, Coudé Adam, Holmström, Stålhammar
Cancers, volume 18, issue 5, article 818. Online 3 March 2026.
doi.org/10.3390/cancers18050818, opens in a new tab Read the paper, Cancers, 2026, opens in a new tab
PubMed
41827752
PubMed Central
PMC12984398
Funding
“This research was funded by Karolinska Institutet, grant number 2023-02622.”
Ethics
Swedish Ethical Review Authority, reference 2025-05054-01. Patient consent was waived because the study was retrospective.
Conflicts
“Mats Holmström and Gustav Stålhammar are founders of Eyedentity AB and hold equity in the company, which develops the MelAInoma software evaluated in this study. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.”

The short version, in nine lines.

Developed, compared, and tested where referrals begin.

Full results, Ophthalmology Science, 2024Sensitivity and specificity for each group of ophthalmologists with P values, the two checklists, and the two camera types, and the terms used.

The readers and the checklists, Ophthalmology Science, 2024Who read the photographs, what MOLES and TFSOM-UHHD score, and what each method had to go on.

Every rule on the same images, TVST, 2025The guideline's MOLES rule and a stricter cut-off against the model's threshold, what “quadrupled” and “10‑fold” compare, net benefit, and the spread between opticians.

After trying the tool, TVST, 2025What the 25 opticians who ran the model on the same images said afterwards, and how the question about time was worded.

MOLES, letter by letter, TVST, 2025The five signs, how they are scored, and the Swedish referral guideline the opticians followed.

Full results, Cancers, 2026Odds ratios, explained variation, how much the two scores overlap, and the two scores combined, and the terms used.

The 86 lesions, Cancers, 2026Sizes, cameras, how the MOLES scores were spread, and how each lesion's reference answer was reached.

The model finds the lesion first, then scores it.

Find the lesion

Crop around it

Score from 0 to 1

Every label came from a specialist's examination.

How each image got its answer

One question, asked of one photograph.

Why earlier matters, in the group's own researchSize, delay and the limits of the growth checklists, from Prof. Stålhammar's wider research, with his own caveats.

What the studies leave open.

The limitations, paper by paperThe limits the three papers list, in the authors' words, grouped by paper.

The people behind the evidence.

Prof. Gustav Stålhammar

Authors on the three papers

Selected wider research by Prof. Stålhammar and colleaguesTen papers on the disease itself: how often it occurs and how it presents, tumour size and staging, small tumours and their treatment, growth and treatment delay, and earlier work on networks that read pathology slides.

Read the papers in full.

Development and Validation of a Deep Learning Algorithm for Differentiation of Choroidal Nevi from Small Melanoma in Fundus Photographs

How to cite, Ophthalmology Science, 2024

Challenges for Opticians in Evaluating Small Pigmented Choroidal Lesions: Potential Support From the MelAInoma Deep Learning Algorithm

How to cite, TVST, 2025

Comparison of MOLES and MelAInoma for Differentiating Small Choroidal Melanomas from Nevi

How to cite, Cancers, 2026

If you have questions about the studies, their data or their limits, we will go through them with you.