A
recent retrospective cross-sectional observational study concluded that the
Medios AI system, in conjunction with the Remidio smartphone-based fundus
camera, has strong potential to detect common retinal pathologies, such as
diabetic retinopathy (DR), age-related macular degeneration (ARMD), and
glaucoma, in primary or screening settings.
However,
the same study identified potential limitations of the system, including a
tendency to misclassify leukemic retinopathy as diabetic retinopathy, a failure
to detect optic atrophy, and an inability to capture peripheral retinal lesions
due to hardware constraints.
This
retrospective cross-sectional observational study was published in December
2025 in Cureus.
Introduction
Artificial
intelligence (AI) is revolutionizing medical diagnostics by introducing tools
that promise faster, more accessible, and often cost-effective solutions for
disease detection and treatment. It has recently been integrated into
ophthalmic diagnostics for screening retinal diseases. The Medios AI system by
Remidio (Singapore, Singapore) detects diabetic retinopathy (DR), age-related
macular degeneration (ARMD), and glaucoma; however, its performance in ocular
oncology remains underexplored. Ocular oncology involves intraocular tumors,
paraneoplastic syndromes, and systemic malignancies with ocular manifestations.
Retinal findings can include leukemic retinopathy, choroidal metastases, optic
nerve atrophy, and hemorrhagic retinopathy, which may present subtly or
overlap.
Study Overview
A retrospective cross-sectional study was
conducted at the National Cancer Institute, All India Institute of Medical
Sciences, New Delhi, to assess the diagnostic performance of the Medios AI
system in patients with ocular cancer. The study enrolled 98 patients (196
eyes) with systemic malignancies, including those undergoing treatment for
leukaemia, lymphoma, and other tumours, who presented with ocular complaints.
Patients with corneal ulcers, significant cataract, vitreous haemorrhage
preventing fundus imaging, and those who were uncooperative were excluded.
Fundus images were captured with a smartphone-based camera (Remidio) under
standard lighting. Both the macular and optic disc regions were imaged, with
peripheral retina when possible. The primary outcome was diagnostic concordance
between the AI-generated findings and clinical diagnoses. Secondary outcomes
included false positives, false negatives, misclassification analysis, and
imaging system limitations.
Key
Findings
Diagnostic performance of the Medios AI system
The AI identified glaucomatous cupping in
three patients, confirmed by clinical exam, visual field tests, and OCT,
showing optic disc ratios >0.7 and rim thinning. It flagged 10 cases of
diabetic retinopathy, but only 2 were confirmed; 8 patients had leukemic
retinopathy with similar retinal features, suggesting low specificity. The AI
detected early-to-intermediate dry ARMD in two patients, with drusen and RPE
irregularities. It misclassified 8 cases of leukemic retinopathy as DR,
indicating difficulty distinguishing similar vascular retinopathies. The AI
missed optic atrophy in 5 patients despite clinical signs of optic disc pallor,
which is concerning given the importance of optic nerve issues in oncology
patients with CNS complications (Table 1).
Table 1: Diagnostic performance of Medios
AI system
Retinal Condition
Cases Detected by AI
Confirmed Clinically
Correct Diagnoses
Misclassified / Missed Findings
% Missed
Glaucoma (Cupping)
3
3
3
None
0%
Diabetic Retinopathy
10
2
2
8 cases of leukemic retinopathy
misclassified
80%
Age-Related Macular Degeneration (ARMD)
2
2
2
None
0%
Leukemic Retinopathy
0 (all misclassified)
8
0
All 8 labelled as diabetic retinopathy by
AI
100%
Optic Atrophy
0
5
0
5 completely missed
100%
Peripheral Retinal Lesions (e.g.,
metastases, infiltrates)
Not detectable (hardware limitation)
Several noted clinically
0
Missed due to lack of wide-field/montage
imaging
Not applicable
Clinical
Implications
Medios
AI system shows potential for identifying prevalent retinal pathologies, such
as diabetic retinopathy (DR), age-related macular degeneration (ARMD), and
glaucoma; however, its limitations are pronounced in oncology. The system’s
inability to differentiate between similar haemorrhagic retinopathies, failure
to detect optic nerve atrophy, and lack of wide-field imaging capabilities
highlight the need for cautious implementation. However, the integration of AI
tools must be accompanied by expert clinical oversight, particularly in specialised
settings where retinal presentations are complex and atypical.
Reference: Das D, Chawla B,
Lomi N, et al. AI in the Shadows: Unveiling the Strengths and Blind Spots of
Medios AI Retinal Screening in Cancer Care. Cureus 17(12): e99002. Published December
11, 2025. DOI 10.7759/cureus.99002