Nakod69/qai-oli-models

🏛️ QAI-OLI: Model Checkpoints & Access Policy

Tri-Stage Quantum-Classical Oral Oncology Cascade (83.77% SOTA | 98.40% Screening Triage)

---

🔐 Model Weights Access Policy

To prevent unauthorized distribution and uphold clinical data governance, the trained PyTorch model checkpoints (.pth binary weight files) are not hosted publicly in this GitHub repository.

📧 How to Request Model Checkpoints:

If you are an academic researcher, clinical practitioner, or institutional collaborator seeking model weights for validation or benchmark reproduction, please contact the primary developer:

• Primary Contact / Lead Developer: Kshitij Nakod
• Email: nakod.kshitij5228@gmail.com
• GitHub: https://github.com/kshitij5228
• Email Subject: [QAI-OLI] Model Weights Access Request
• Please include: Your full name, affiliated institution/university, and intended research/clinical use case.

👥 Project Authors & Research Engineering Team (ARMIET, University of Mumbai):

1. Kshitij Nakod (Quantum AI & Model Lead) — GitHub: @kshitij5228 | Email: nakod.kshitij5228@gmail.com
2. Amit Chauhan (Model Fine-Tuning & Data) — GitHub: @amitchn28
3. Manoj Yadav (Full-Stack & Web Systems) — GitHub: @Spmanoj593
4. Sahil Khadkar (Explainable AI & Testing) — GitHub: @sahilkhadkar2005-bit

---

📁 Required Checkpoint Files

Once access is granted, download and place the three .pth checkpoint files directly into this models/ directory:

Filename	Architecture	Size	Specialized Clinical Role	Standalone Benchmark
01_mahanet_convnextv2_tiny.pth	ConvNeXt-V2-Tiny	107.1 MB	Spatial Vision & Border Specialist (Captures irregular ulcer edges and macroscopic lesion margins)	70.06% Acc (83.2% OCA F1)
02_mahanet_densenet121.pth	DenseNet-121	28.1 MB	Cellular Morphology & Malignancy Specialist (Dense gradient flow for epithelial atypia and vascular patterns)	61.60% Acc (80.5% OCA F1)
03_mahanet_quantum_8q_vqc.pth	8-Qubit Parameterized VQC Hybrid	152.1 MB	Quantum Hilbert Space Decider (PennyLane 256-D state rotation resolving Benign vs. OPMD ambiguity)	78.64% Acc (84.9% OCA F1)

---

🔄 Parallel Tri-Stream Clinical Ensemble Pipeline (server.py)

At inference time, the system executes all three model streams simultaneously in parallel, combining their calibrated softmax predictions via weighted soft-voting to produce a unified 4-class diagnosis:

1. Stream 1 (Stage 1): ConvNeXt-V2-Tiny (01_mahanet_convnextv2_tiny.pth) — Weight: 25%

   • Role: Extracts macro lesion contours, irregular borders, and structural shape representations (70.06% Standalone Acc).

2. Stream 2 (Stage 2): DenseNet-121 (02_mahanet_densenet121.pth) — Weight: 25%

   • Role: Captures fine epithelial micro-vascular patterns and cellular atypia via dense iterative feature reuse (61.60% Standalone Acc).

3. Stream 3 (Stage 3): 8-Qubit Parameterized VQC Hybrid (03_mahanet_quantum_8q_vqc.pth) — Weight: 50%

   • Role: Projects deep latent features into a 256-D quantum Hilbert space via parameterized AngleEmbedding + Entangling layers, with a 64-D classical skip highway (78.64% Standalone Acc).

🎯 Consensus Soft-Voting Formulation:

$${\mathbf{P}}_{\text{fused}}=0.25\cdot {\mathbf{P}}_{\text{M1}}+0.25\cdot {\mathbf{P}}_{\text{M2}}+0.50\cdot {\mathbf{P}}_{\text{M3}}\backslash mathbf\{P\}\_\{\backslash text\{fused\}\}\; =\; 0.25\; \backslash cdot\; \backslash mathbf\{P\}\_\{\backslash text\{M1\}\}\; +\; 0.25\; \backslash cdot\; \backslash mathbf\{P\}\_\{\backslash text\{M2\}\}\; +\; 0.50\; \backslash cdot\; \backslash mathbf\{P\}\_\{\backslash text\{M3\}\}$$

🏥 Multi-Tier Clinical Validation:

• Healthy Screening Triage: 98.40% Specificity (near-zero false alarms)
• Malignant Carcinoma Isolation: 89.40% OCA F1-Score (90.43% Sensitivity / Recall)
• Benign vs OPMD Dysplasia Decoupling: 83.77% Discrimination Accuracy (resolves tissue mimicry)