Upload MedVisionAI best model (epoch 70, eval_accuracy: 0.9012)

README.md

@@ -0,0 +1,111 @@
+---
+license: apache-2.0
+library_name: timm
+---
+# MedVisionAI
+<!-- markdownlint-disable first-line-h1 -->
+<!-- markdownlint-disable html -->
+<!-- markdownlint-disable no-duplicate-header -->
+
+<div align="center">
+  <img src="figures/fig1.png" width="60%" alt="MedVisionAI" />
+</div>
+<hr>
+
+<div align="center" style="line-height: 1;">
+  <a href="LICENSE" style="margin: 2px;">
+    <img alt="License" src="figures/fig2.png" style="display: inline-block; vertical-align: middle;"/>
+  </a>
+</div>
+
+## 1. Introduction
+
+MedVisionAI represents a breakthrough in medical imaging analysis. This latest version incorporates state-of-the-art vision transformer architectures with domain-specific pre-training on over 15 million anonymized medical images. The model excels in detecting abnormalities across multiple imaging modalities including X-rays, CT scans, MRI, ultrasound, and pathology slides.
+
+<p align="center">
+  <img width="80%" src="figures/fig3.png">
+</p>
+
+Compared to the previous version, MedVisionAI v2 shows remarkable improvements in sensitivity and specificity metrics. For instance, in the ChestX-ray14 benchmark, our model achieves 94.2% AUC compared to 88.7% in the previous version. This advancement comes from our novel multi-scale attention mechanism and curriculum learning strategy during pre-training.
+
+Beyond improved detection accuracy, this version also offers faster inference times (3x speedup) and reduced false positive rates across all imaging modalities.
+
+## 2. Evaluation Results
+
+### Comprehensive Benchmark Results
+
+<div align="center">
+
+| | Benchmark | ResNet-152 | EfficientNet-B7 | ViT-Large | MedVisionAI |
+|---|---|---|---|---|---|
+| **Radiology Tasks** | X-Ray Detection | 0.821 | 0.845 | 0.867 | 0.872 |
+| | CT Segmentation | 0.756 | 0.778 | 0.792 | 0.839 |
+| | MRI Analysis | 0.702 | 0.731 | 0.758 | 0.840 |
+| **Pathology Tasks** | Pathology Classification | 0.834 | 0.856 | 0.871 | 0.918 |
+| | Retinal Screening | 0.798 | 0.812 | 0.835 | 0.873 |
+| | Skin Lesion | 0.765 | 0.789 | 0.801 | 0.867 |
+| **Oncology Tasks** | Tumor Detection | 0.812 | 0.834 | 0.856 | 0.891 |
+| | Mammography | 0.789 | 0.801 | 0.823 | 0.882 |
+| | Lung Nodule | 0.745 | 0.768 | 0.789 | 0.851 |
+| **Specialty Tasks** | Bone Fracture | 0.867 | 0.878 | 0.891 | 0.921 |
+| | Cardiac Imaging | 0.723 | 0.745 | 0.767 | 0.840 |
+| | Brain Lesion | 0.698 | 0.721 | 0.745 | 0.811 |
+| | Dental Analysis | 0.812 | 0.834 | 0.851 | 0.885 |
+| | Ultrasound | 0.689 | 0.712 | 0.734 | 0.825 |
+| | Anomaly Detection | 0.756 | 0.778 | 0.801 | 0.844 |
+
+</div>
+
+### Overall Performance Summary
+MedVisionAI demonstrates exceptional performance across all evaluated medical imaging benchmark categories, with particularly notable results in oncology detection and pathology classification tasks.
+
+## 3. Clinical Integration & API Platform
+We offer HIPAA-compliant API endpoints for clinical integration. Please contact our enterprise team for deployment options in healthcare settings.
+
+## 4. How to Run Locally
+
+Please refer to our code repository for information about deploying MedVisionAI in your medical imaging pipeline.
+
+Key usage recommendations:
+
+1. Input images should be DICOM format or high-resolution PNG/JPEG.
+2. Pre-processing normalization is handled automatically based on imaging modality.
+
+The model architecture of MedVisionAI-Lite is optimized for edge deployment on medical devices with limited compute resources.
+
+### Image Pre-processing
+We recommend using the following pre-processing pipeline:
+```python
+from medvisionai import MedVisionPreprocessor
+preprocessor = MedVisionPreprocessor(
+    modality="xray",  # or "ct", "mri", "ultrasound", etc.
+    normalize=True,
+    target_size=(512, 512)
+)
+processed_image = preprocessor(raw_dicom)
+```
+
+### Inference Configuration
+We recommend the following confidence thresholds:
+- Screening applications: threshold = 0.3 (high sensitivity)
+- Diagnostic support: threshold = 0.5 (balanced)
+- Triage applications: threshold = 0.7 (high specificity)
+
+### Batch Processing
+For processing multiple images efficiently:
+```python
+from medvisionai import BatchProcessor
+
+batch_processor = BatchProcessor(
+    model_path="./checkpoints/epoch_100",
+    batch_size=32,
+    device="cuda"
+)
+results = batch_processor.process_folder("./patient_scans/")
+```
+
+## 5. License
+This code repository is licensed under the [Apache 2.0 License](LICENSE). The use of MedVisionAI models for clinical applications requires additional certification and compliance verification.
+
+## 6. Contact
+If you have questions about clinical deployment or research collaboration, please contact us at medical@medvisionai.health.

config.json

@@ -0,0 +1,9 @@
+{
+    "model_type": "vit",
+    "architectures": ["ViTForImageClassification"],
+    "eval_accuracy": 0.9012,
+    "epoch": 70,
+    "image_size": 512,
+    "hidden_size": 768,
+    "num_attention_heads": 12
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa05fcbb762847b2374743942911d15466d8d0f08adc0487c26b5b540a268fe8
+size 4004