Upload MedVisionAI best checkpoint (epoch_50) with evaluation results

README.md

@@ -1,6 +1,6 @@
 ---
 license: apache-2.0
-library_name: transformers
+library_name: timm
 ---
 # MedVisionAI
 <!-- markdownlint-disable first-line-h1 -->
@@ -20,15 +20,15 @@ library_name: transformers
 
 ## 1. Introduction
 
-MedVisionAI represents a breakthrough in medical imaging analysis. In this latest version, MedVisionAI has achieved significant improvements in diagnostic accuracy and multi-modal imaging interpretation through advanced transfer learning and domain-specific fine-tuning. The model demonstrates exceptional performance across radiology, pathology, and dermatology imaging benchmarks, approaching human-expert level performance in many categories.
+MedVisionAI represents a breakthrough in medical imaging analysis. In this latest version, MedVisionAI has dramatically enhanced its diagnostic accuracy and multi-modality support by incorporating advanced attention mechanisms and leveraging large-scale clinical imaging datasets during pre-training. The model demonstrates exceptional performance across various medical imaging benchmarks, including X-ray analysis, CT segmentation, and pathology grading.
 
 <p align="center">
   <img width="80%" src="figures/fig3.png">
 </p>
 
-Compared to the previous version, the upgraded model shows remarkable improvements in detecting subtle abnormalities. For instance, in the RadBench-2025 evaluation, the model's tumor detection sensitivity increased from 82% in the previous version to 94.3% in the current version. This advancement stems from enhanced attention mechanisms during the diagnostic reasoning process: the previous model processed images in 2K token context windows, whereas the new version utilizes 8K token context for comprehensive analysis.
+Compared to the previous version, this upgrade shows remarkable improvements in detecting subtle abnormalities. For example, in the RSNA Pneumonia Detection Challenge, the model's sensitivity increased from 78% in the previous version to 91.3% in the current version. This advancement stems from improved feature extraction: the previous model processed images at 224x224 resolution, whereas the new version operates at 512x512 resolution with hierarchical feature fusion.
 
-Beyond its improved diagnostic capabilities, this version also offers reduced false-positive rates and enhanced support for multi-modal inputs combining imaging with clinical notes.
+Beyond diagnostic accuracy improvements, this version also offers enhanced explainability through attention maps and reduced false positive rates in screening applications.
 
 ## 2. Evaluation Results
 
@@ -36,70 +36,70 @@ Beyond its improved diagnostic capabilities, this version also offers reduced fa
 
 <div align="center">
 
-| | Benchmark | RadNet | PathAI | RadNet-v2 | MedVisionAI |
+| | Benchmark | ModelA | ModelB | ModelA-v2 | MedVisionAI |
 |---|---|---|---|---|---|
-| **Core Detection Tasks** | Tumor Detection | 0.823 | 0.845 | 0.861 | 0.818 |
-| | Differential Diagnosis | 0.756 | 0.771 | 0.785 | 0.834 |
-| | Clinical Correlation | 0.698 | 0.712 | 0.725 | 0.733 |
-| **Image Analysis** | Report Interpretation | 0.645 | 0.662 | 0.678 | 0.690 |
-| | Symptom Extraction | 0.712 | 0.729 | 0.745 | 0.737 |
-| | Abnormality Classification | 0.834 | 0.851 | 0.867 | 0.871 |
-| | Diagnostic Confidence | 0.789 | 0.802 | 0.815 | 0.820 |
-| **Generation Tasks** | Image Segmentation | 0.678 | 0.695 | 0.712 | 0.750 |
-| | Case Documentation | 0.623 | 0.641 | 0.658 | 0.623 |
-| | Patient Communication | 0.701 | 0.718 | 0.735 | 0.649 |
-| | Findings Summary | 0.756 | 0.773 | 0.789 | 0.770 |
-| **Specialized Capabilities**| Multilingual Reports | 0.812 | 0.829 | 0.845 | 0.847 |
-| | Medical Knowledge | 0.734 | 0.751 | 0.767 | 0.712 |
-| | Protocol Adherence | 0.778 | 0.795 | 0.812 | 0.809 |
-| | Patient Safety | 0.892 | 0.908 | 0.923 | 0.893 |
+| **Diagnostic Imaging** | X-Ray Classification | 0.821 | 0.835 | 0.842 | 0.800 |
+| | Tumor Detection | 0.756 | 0.771 | 0.783 | 0.769 |
+| | CT Segmentation | 0.689 | 0.705 | 0.718 | 0.796 |
+| **Specialized Analysis** | MRI Analysis | 0.734 | 0.749 | 0.761 | 0.750 |
+| | Pathology Grading | 0.692 | 0.708 | 0.721 | 0.678 |
+| | Retinal Screening | 0.803 | 0.819 | 0.828 | 0.864 |
+| | Bone Fracture | 0.768 | 0.782 | 0.795 | 0.783 |
+| **Detection Tasks** | Ultrasound Detection | 0.645 | 0.661 | 0.673 | 0.640 |
+| | Skin Lesion | 0.712 | 0.728 | 0.741 | 0.695 |
+| | Mammography | 0.778 | 0.793 | 0.805 | 0.842 |
+| | Organ Localization | 0.701 | 0.718 | 0.729 | 0.769 |
+| **Advanced Capabilities** | Cardiac Imaging | 0.723 | 0.738 | 0.751 | 0.685 |
+| | Brain Anomaly | 0.681 | 0.697 | 0.709 | 0.665 |
+| | Lung Nodule | 0.745 | 0.761 | 0.773 | 0.782 |
+| | Safety Compliance | 0.812 | 0.801 | 0.825 | 0.800 |
 
 </div>
 
 ### Overall Performance Summary
-MedVisionAI demonstrates exceptional performance across all evaluated medical imaging benchmark categories, with particularly notable results in detection and safety-critical tasks.
+MedVisionAI demonstrates strong performance across all evaluated medical imaging benchmark categories, with particularly notable results in diagnostic imaging and specialized analysis tasks.
 
 ## 3. Clinical Integration & API Platform
-We offer HIPAA-compliant API endpoints and clinical integration services. Please check our official website for compliance documentation and integration guides.
+We offer a clinical integration interface and API for healthcare providers to integrate MedVisionAI. Please check our official documentation for HIPAA-compliant deployment details.
 
 ## 4. How to Run Locally
 
-Please refer to our clinical deployment repository for information about running MedVisionAI in healthcare environments.
+Please refer to our code repository for more information about running MedVisionAI locally.
 
-Compared to previous versions, the deployment recommendations for MedVisionAI have the following changes:
+Compared to previous versions, the usage recommendations for MedVisionAI have the following changes:
 
-1. GPU with minimum 24GB VRAM is recommended for optimal inference speed.
-2. Multi-image batch processing is now supported for radiology workflows.
+1. DICOM input format is now natively supported.
+2. Multi-GPU inference is available for high-resolution 3D volume processing.
 
-The model architecture of MedVisionAI-Lite is optimized for edge deployment while maintaining diagnostic accuracy above 95% of the full model.
+The model architecture of MedVisionAI-Lite is optimized for edge deployment, sharing the same preprocessing pipeline as the main MedVisionAI model.
 
-### System Configuration
-We recommend using the following clinical context prompt:
-```
-You are MedVisionAI, an AI assistant for medical imaging analysis.
-Current examination date: {exam_date}
-Patient context: {patient_context}
+### Preprocessing Configuration
+We recommend using the following preprocessing configuration for optimal results.
+```python
+preprocessing_config = {
+    "target_spacing": [1.0, 1.0, 1.0],  # mm
+    "intensity_normalization": "z-score",
+    "window_center": 40,
+    "window_width": 400
+}
 ```
 
-### Confidence Thresholds
-We recommend setting the diagnostic confidence threshold to 0.85 for clinical alerts.
+### Inference Settings
+We recommend setting the confidence threshold $\tau$ to 0.7 for clinical screening applications.
 
-### Input Formats
-For DICOM image analysis, use the following template:
+### Input Format Requirements
+For DICOM input, please ensure the following metadata fields are present:
 ```
-image_template = \
-"""[modality]: {modality}
-[body_region]: {body_region}
-[clinical_indication]: {indication}
-[image_data_begin]
-{encoded_image}
-[image_data_end]
-{diagnostic_question}"""
+Required DICOM Tags:
+- PatientID (anonymized)
+- StudyDate
+- Modality
+- PixelSpacing
+- SliceThickness (for 3D volumes)
 ```
 
 ## 5. License
-This code repository is licensed under the [Apache 2.0 License](LICENSE). The use of MedVisionAI models is subject to additional healthcare compliance requirements detailed in our clinical deployment guide.
+This code repository is licensed under the [Apache License 2.0](LICENSE). The use of MedVisionAI models is subject to additional clinical validation requirements. The model supports research use and requires FDA clearance for clinical deployment.
 
 ## 6. Contact
-For clinical partnerships and integration support, contact us at clinical@medvisionai.health
-For research collaborations: research@medvisionai.health
+If you have any questions, please raise an issue on our GitHub repository or contact us at support@medvisionai.health.

config.json

@@ -1,13 +1,4 @@
 {
     "model_type": "vit",
-    "architectures": [
-        "ViTForImageClassification"
-    ],
-    "hidden_size": 768,
-    "num_attention_heads": 12,
-    "num_hidden_layers": 12,
-    "image_size": 512,
-    "patch_size": 16,
-    "num_channels": 3,
-    "num_labels": 14
-}
+    "architectures": ["ViTForImageClassification"]
+  }