Editing AI for Ophthalmology

<div style="background-color: #4B0082; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">
{{BloomIntro}}
AI for ophthalmology applies deep learning to the diagnosis and monitoring of eye diseases using retinal images, OCT scans, fundus photographs, and visual field data. The eye is uniquely accessible for imaging — retinal photographs can be taken non-invasively in minutes. AI systems trained on millions of retinal images can detect diabetic retinopathy, glaucoma, age-related macular degeneration, and rare retinal conditions with ophthalmologist-level accuracy. Google's 2016 demonstration of AI achieving expert-level diabetic retinopathy screening was a landmark moment for medical AI, and the field has advanced dramatically since.
</div>

__TOC__

<div style="background-color: #000080; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">
== <span style="color: #FFFFFF;">Remembering</span> ==
* '''Diabetic retinopathy (DR)''' — Damage to retinal blood vessels from diabetes; leading cause of blindness in working-age adults; highly amenable to AI screening.
* '''Glaucoma''' — Optic nerve damage causing visual field loss; detected by AI from optic disc analysis and OCT nerve fiber layer thickness.
* '''Age-related macular degeneration (AMD)''' — Degeneration of the macula (central retina); leading cause of blindness in elderly; AI detects and grades severity.
* '''Fundus photograph''' — Non-invasive color photo of the retina; primary modality for DR and AMD screening.
* '''OCT (Optical Coherence Tomography)''' — High-resolution cross-sectional imaging of retinal layers; essential for AMD and glaucoma assessment.
* '''International Clinical DR scale''' — Standardized severity grading (0–4) for DR; AI automates this grading.
* '''IDx-DR''' — First FDA-authorized AI diagnostic device (2018); autonomously detects more-than-mild DR from fundus photos.
* '''Google DeepMind (retinal AI)''' — Foundational 2016 paper demonstrating AI matching ophthalmologist DR grading accuracy.
* '''Diabetic macular edema (DME)''' — Fluid accumulation in the macula from diabetic retinopathy; key finding requiring treatment.
* '''VEGF (Vascular Endothelial Growth Factor)''' — Target of anti-VEGF injections for AMD and DME; AI predicts treatment response.
* '''Optic disc''' — The point where the optic nerve enters the retina; key feature for glaucoma assessment.
* '''Cup-to-disc ratio (CDR)''' — Ratio of optic cup to optic disc diameter; elevated CDR is a glaucoma risk indicator.
* '''Choroidal thickness''' — Measured by OCT; biomarker for multiple retinal diseases.
* '''EyePACS''' — A large-scale retinal image dataset used to train DR screening AI; >1M images.
</div>

<div style="background-color: #006400; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">
== <span style="color: #FFFFFF;">Understanding</span> ==
Ophthalmology AI is one of the clearest demonstrations of AI's potential in medicine: a well-defined screening task (does this patient have sight-threatening DR?), a scalable imaging modality (fundus photography takes seconds), a clear clinical need (most people with diabetes never receive recommended annual screening), and a validated gold standard (ophthalmologist grading).

**Google's 2016 breakthrough**: Gulshan et al. (2016) trained a deep CNN on 128,175 retinal images graded by 54 ophthalmologists. The AI achieved AUC 0.99 for detecting referrable DR — matching the performance of expert ophthalmologists. This paper, published in JAMA, was a pivotal moment demonstrating that AI could match specialist performance on a clinically significant task.

**From research to reality (IDx-DR)**: In 2018, FDA authorized IDx-DR as the first AI autonomous diagnostic device — meaning a doctor need not interpret the results. Primary care physicians can use it to screen patients for DR without ophthalmologist involvement. A positive result triggers referral; negative clears the patient for 12 months. This is transformative for settings without access to ophthalmologists.

**Beyond DR — multi-disease screening**: AI systems now screen for multiple conditions simultaneously from the same fundus photograph. DeepMind's 2018 Nature Medicine paper showed AI detecting 50+ ophthalmic conditions from OCT scans with specialist-level accuracy, predicting need for urgent referral across diverse disease categories.

**Retinal biomarkers for systemic disease**: Remarkably, retinal images contain information about systemic health beyond the eye. Google's 2018 paper showed AI predicting age, sex, blood pressure, smoking status, and cardiovascular risk directly from fundus photos — features not previously known to be visible to human clinicians. This opens a new dimension of retinal AI: using the eye as a window to systemic health.
</div>

<div style="background-color: #8B0000; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">
== <span style="color: #FFFFFF;">Applying</span> ==
'''Diabetic retinopathy grading from fundus images:'''
<syntaxhighlight lang="python">
import torch
import torch.nn as nn
from torchvision import models, transforms
from torch.utils.data import Dataset, DataLoader
from PIL import Image
import pandas as pd
import numpy as np
from sklearn.metrics import cohen_kappa_score, roc_auc_score

class RetinalDataset(Dataset):
    """Retinal fundus image dataset for DR grading."""
    DR_GRADES = {0: 'No DR', 1: 'Mild', 2: 'Moderate', 3: 'Severe', 4: 'Proliferative'}
    def __init__(self, df, transform=None):
        self.df = df
        self.transform = transform

    def __len__(self): return len(self.df)
    def __getitem__(self, idx):
        row = self.df.iloc[idx]
        img = Image.open(row['image_path']).convert('RGB')
        if self.transform: img = self.transform(img)
        return img, int(row['dr_grade'])

# EfficientNet-B4 for DR grading (strong baseline)
transform = transforms.Compose([
    transforms.Resize(380), transforms.CenterCrop(380),
    transforms.RandomHorizontalFlip(), transforms.RandomVerticalFlip(),
    transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

model = models.efficientnet_b4(weights='IMAGENET1K_V1')
model.classifier[1] = nn.Linear(model.classifier[1].in_features, 5)  # 5 DR grades

# Ordinal classification loss (preserves grade ordering)
class OrdinalLoss(nn.Module):
    """Convert ordinal regression to K-1 binary classifiers."""
    def __init__(self, n_classes=5):
        super().__init__()
        self.n_classes = n_classes
    def forward(self, logits, targets):
        cumulative_targets = torch.zeros(len(targets), self.n_classes - 1)
        for i, t in enumerate(targets):
            cumulative_targets[i, :t] = 1  # 1 for all grades ≤ target
        return nn.BCEWithLogitsLoss()(logits[:, :-1], cumulative_targets.to(logits.device))

criterion = OrdinalLoss(n_classes=5)
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4, weight_decay=1e-2)

# Key metric: quadratic weighted kappa (Kaggle DR competition metric)
def quadratic_kappa(y_true, y_pred):
    return cohen_kappa_score(y_true, y_pred, weights='quadratic')
</syntaxhighlight>

; Ophthalmology AI tools
: '''Commercial DR screening''' → IDx-DR (Digital Diagnostics), Eyenuk EyeArt, Retinalyze
: '''Multi-disease screening''' → DeepMind Streams (now Verily), Notal Vision
: '''AMD/OCT analysis''' → Heidelberg HEYEX AI, Zeiss CIRRUS AI
: '''Glaucoma''' → Glaukos AI, Ora ClinicalEdge, Optomed Aurora
: '''Datasets''' → EyePACS (1M+ images), Kaggle DR, DRIVE, ORIGA, REFUGE
</div>

<div style="background-color: #8B4500; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">
== <span style="color: #FFFFFF;">Analyzing</span> ==
{| class="wikitable"
|+ Ophthalmology AI Performance Summary
! Condition !! Modality !! Best AI AUC !! vs. Ophthalmologist
|-
| Diabetic retinopathy || Fundus || 0.99 || Matches/exceeds
|-
| AMD (referable) || OCT || 0.99 || Matches
|-
| Glaucoma || Fundus+OCT || 0.95 || Matches
|-
| Retinopathy of prematurity || Fundus || 0.98 || Matches
|-
| Diabetic macular edema || OCT || 0.97 || Matches
|-
| Systemic disease biomarkers || Fundus || 0.7-0.85 || Novel (no human baseline)
|}

'''Failure modes''': Image quality sensitivity — poor-quality fundus photos (cataracts, inadequate dilation, motion blur) cause AI failures; quality filters are critical. Demographic generalization — training on predominantly one ethnicity may reduce performance for others (pigmentation affects fundus appearance). Grade disagreement — DR grading itself has significant inter-grader disagreement; training on average labels loses boundary cases. Autonomous AI replacing referral vs. assisting screening — different regulatory and workflow implications.
</div>

<div style="background-color: #483D8B; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">
== <span style="color: #FFFFFF;">Evaluating</span> ==
Ophthalmology AI evaluation: (1) **AUC per condition**: DR, AMD, glaucoma separately. (2) **Sensitivity/specificity at clinical threshold**: referral sensitivity >90% for DR screening (missing DR is worse than unnecessary referral). (3) **Image quality gating**: measure performance on images that pass quality filter vs. all images. (4) **Multi-site validation**: different scanner types (Topcon, Canon, Zeiss, Optomed) and patient populations. (5) **Reader study**: ophthalmologists with and without AI assistance — does AI add value beyond ophthalmologist alone?
</div>

<div style="background-color: #2F4F4F; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">
== <span style="color: #FFFFFF;">Creating</span> ==
Deploying an AI DR screening program: (1) Camera: portable non-mydriatic fundus camera (Topcon NW400, Optomed Aurora) for primary care screening. (2) Image quality: automatic quality check (image brightness, focus, disc visibility) before AI analysis. (3) AI grading: EfficientNet-B4 fine-tuned on ≥10,000 graded images; threshold for referral = grade ≥ 2 (moderate DR). (4) Integration: connect to EMR; flag patients due for annual screening; automated results letter to patient. (5) Referral pathway: positive AI → ophthalmology referral with image; negative AI → 12-month recall. (6) Quality assurance: quarterly audit of 50 cases by ophthalmologist; recalibrate threshold if performance drifts.

[[Category:Artificial Intelligence]]
[[Category:Ophthalmology]]
[[Category:Medical Imaging]]
</div>