Eye sample

The samples and clinical features

Both original 1728 × 1296 photographs are embedded unchanged. The cataract and normal labels were supplied by the user; no clinician report, cataract subtype, or severity accompanies these files. Pupil/reflection boundaries and crop placement are manual guides. Visible cloudiness in this example is not evidence of early-cataract performance. LOCS III uses slit-lamp standards for nuclear color/opalescence and retroillumination standards for cortical and posterior subcapsular opacity. DeepLensNet studied related clinical grading with distinct slit-lamp and retroillumination views. Those protocols do not establish which small lesions CVION can resolve.

The hardware and capture

The 3D camera is the approved IMX708 sensor-and-lens reference assembly, without the full Raspberry Pi board. The reference sensor has 4608 × 2592 pixels at 1.4 µm pitch and phase-detection autofocus. The depicted housing/optics use the standard sensor-assembly reference; the exact Camera Module 3 variant and additional optics still need confirmation. Current code selects a 1728 × 1296 crop from a 4608 × 2592, 10-bit sensor mode and saves JPEG at quality 95. These are software settings, not verified acquisition metadata for the supplied sample. Pixel pitch and pixel count do not establish object-space resolution, focus quality, or early-cataract sensitivity.

The illumination matches the reported architecture: 3 W star LED, collimating lens, round aperture, approximately 45° to the camera axis. Model dimensions, mounting distances, beam footprint, and light paths are illustrative. The camera-facing eye card is a photograph, not a volumetric eye or tissue-scattering simulation. The round opening limits the accepted bundle; a finite LED still produces residual divergence. It does not form the thin optical section of a slit lamp. Optical output, irradiance, exposure, beam diameter at the eye, and glare reduction have not been measured here. The 3 W rating alone does not establish them. The cutaway separates components only to explain image formation. The image emerging from it is the supplied JPEG, not a new simulated capture.

The proposed inference workflow

The approved full-page pipeline is integrated here without an outer flow diagram: 0. Captured eye image → 1. Quality check → 2. Pupil localization → 3. Classifier model → 4. Results. Framing, focus, and exposure/reflection sweeps illustrate review activity; they do not compute a quality score or pass/fail. A manually placed 288 × 288 pupil crop at (499, 411) is resized to an actual 224 × 224 RGB input using OpenCV INTER_AREA. Automatic pupil localization remains proposed.

The moving window, local-pattern grids, and eight feature-vector slots are schematic model internals, not computed learned activations. The slots illustrate a pooled feature vector; they are not logits. Feature packets feed a classification head before separate class outputs emerge as two muted bars labeled Class A and Class B, placeholders for the two target categories. The model frame groups Local patterns, Feature vector, and Classification head; the cropped image remains outside it as input. The bar caption explains the score comparison without implying a calibrated probability. The head symbol includes scoring and class-value conversion schematically; its node count does not specify a model architecture. Fixed bar heights 0.72 and 0.28 are illustrative values, not probabilities, predictions, or measured performance for this eye. No trained model runs here. A candidate implementation is prepared RGB → pretrained ResNet18 → 512 pooled features → a clinician-labeled logistic-regression head, requiring consistent training/inference preparation and clinical evaluation. The earlier detailed pixel/filter chapter remains available separately.

The experiment that would establish usefulness

Define an assessable capture with an ophthalmologist; record the camera/optics, working distance, focus, exposure, LED setting, and beam size. Compare repeat captures under controlled illumination with a documented clinical reference, including mild cases under a defined early-cataract criterion. Train with per-eye clinical labels and split by patient, keeping repeat images and visits together. Compare the proposed transfer-learning route against the current HOG/SVM research baseline and report held-out sensitivity, specificity, uncertainty, and the rate of captures that cannot be assessed. The current prototype code pools both eyes across a session; that is a different task and cannot establish per-eye performance.