iejfoiesf/adaptive-crossattention-oct-classifier

Adaptive Cross-Attention Multimodal OCT Classifier

Patent-pending · Provisional Specification Filed (India, Feb 2026)

Interpretable, uncertainty-aware retinal disease diagnosis from 2D OCT B-scans using bidirectional cross-attention fusion of visual and vascular-geometric features.

---

Architecture Overview

Stage	Module	Output
Preprocessing	CLAHE + Sato tubeness	float32 [0,1] 512×512
Vessel segmentation	SoftUNet (Dice+BCE)	Prob map 512×512
Geometric encoder	8 metrics · median-agg · CorrectionMLP	g ∈ ℝ⁶⁴
Visual backbone	EfficientNetV2-S (timm)	F ∈ ℝ¹²⁸⁰×⁷×⁷
CAFM	Bidirectional cross-attention (geo↔vis)	z ∈ ℝ⁵¹²
Classifier	MLP + MC Dropout T=20 + Temp Scaling	probs + σ
XAI	GradCAM + Otsu adaptive threshold	saliency PNG
Rationale	Z-score → phrase lookup	clinical string

Classes: CNV · DME · DRUSEN · NORMAL
Dataset: Kermany OCT 2017 (84,495 images)

---

Quick Start

pip install -r requirements.txt

# 1. Train UNet (provide vessel mask labels)
python -m train.precompute_prob_maps --data_dir data/kermany --output_dir prob_maps

# 2. Train classifier
python -m train.train --data_dir data/kermany --prob_maps_dir prob_maps --output_dir weights

# 3. Launch API
python run_api.py

API Endpoints

Method	Endpoint	Description
GET	/health	Service health + device
GET	/classes	Class list
GET	/metrics/normative	Normative μ ± σ for z-scores
POST	/predict	Single OCT image → full JSON
POST	/predict/batch	Up to 16 images

Response includes: predicted class, calibrated confidence, epistemic uncertainty, 8 vascular metrics, normative z-scores, GradCAM saliency (base64 PNG), clinical rationale.

Novel Contributions (Patent Claims)

1. Bidirectional CAFM — mutual geo↔vis cross-attention for 2D OCT fusion
2. Segmentation-Robust Geometric Encoder — multi-threshold median aggregation + residual CorrectionMLP
3. Adaptive-Threshold GradCAM — Otsu thresholding on normalised CAM (replaces fixed τ=0.5)
4. MC Dropout + Temperature Calibration — calibrated confidence + epistemic uncertainty
5. Vascular-Feature-Linked Clinical Rationale — z-score → phrase lookup by (class, metric, direction)

Citation

@misc{oct_cafm_2026,
  title  = {Adaptive Cross-Attention Multimodal OCT Classifier},
  year   = {2026},
  note   = {Patent pending (India provisional)},
  url    = {https://github.com/ARNiteshKumar/adaptive-crossattention-oct-classifier}
}

License

Apache 2.0 — see LICENSE.