Upload MedVisionNet best checkpoint (epoch_10) with evaluation results

README.md

@@ -20,87 +20,97 @@ library_name: transformers
 
 ## 1. Introduction
 
-MedVisionNet represents a breakthrough in medical imaging AI. This latest release significantly enhances diagnostic accuracy across multiple imaging modalities by leveraging advanced vision transformer architectures and specialized pre-training on diverse medical datasets. The model demonstrates state-of-the-art performance across radiology, pathology, and ophthalmology benchmarks.
+MedVisionNet represents a breakthrough in medical image analysis using vision transformers. This latest version incorporates advanced attention mechanisms specifically designed for radiological image interpretation. The model demonstrates state-of-the-art performance across multiple medical imaging modalities including X-ray, CT, MRI, and ultrasound.
 
 <p align="center">
   <img width="80%" src="figures/fig3.png">
 </p>
 
-Compared to the previous version, MedVisionNet shows remarkable improvements in detecting subtle abnormalities. For instance, in the ChestX-ray14 pneumonia detection task, the model's AUC has improved from 0.82 in the previous version to 0.91 in the current release. This advancement stems from our novel multi-scale attention mechanism specifically designed for medical imaging contexts.
+Compared to the previous version, MedVisionNet v2 shows remarkable improvements in segmentation tasks. In the BraTS 2024 challenge, the model's mean Dice score improved from 0.82 in the previous version to 0.91 in the current version. This advancement stems from the multi-scale feature pyramid network architecture: the previous model processed images at a single resolution, whereas the new version analyzes at 4 different scales simultaneously.
 
-Beyond improved detection capabilities, this version features reduced false positive rates and enhanced interpretability through attention map visualization.
+Beyond its improved segmentation capabilities, this version also provides uncertainty quantification and enhanced interpretability through attention map visualization.
 
 ## 2. Evaluation Results
 
-### Comprehensive Medical Imaging Benchmark Results
+### Comprehensive Benchmark Results
 
 <div align="center">
 
-| | Benchmark | ResNet-152 | EfficientNet-B7 | ViT-Large | MedVisionNet |
+| | Benchmark | ModelA | ModelB | ModelA-v2 | MedVisionNet |
 |---|---|---|---|---|---|
-| **Radiology Tasks** | Chest X-Ray Classification | 0.823 | 0.845 | 0.861 | 0.818 |
-| | Lung Nodule Detection | 0.756 | 0.778 | 0.792 | 0.800 |
-| | Bone Fracture Detection | 0.812 | 0.831 | 0.847 | 0.859 |
-| **CT/MRI Analysis** | CT Segmentation | 0.721 | 0.743 | 0.761 | 0.700 |
-| | MRI Tumor Detection | 0.789 | 0.812 | 0.829 | 0.885 |
-| | Brain MRI Analysis | 0.734 | 0.756 | 0.778 | 0.753 |
-| | Liver Lesion Detection | 0.698 | 0.721 | 0.739 | 0.691 |
-| **Ophthalmology** | Fundus Grading | 0.845 | 0.867 | 0.881 | 0.841 |
-| | Retinal OCT Analysis | 0.812 | 0.834 | 0.851 | 0.842 |
-| **Dermatology** | Dermoscopy Detection | 0.778 | 0.801 | 0.819 | 0.874 |
-| **Pathology** | Pathology Slides | 0.689 | 0.712 | 0.731 | 0.673 |
-| **Specialized** | Mammography Screening | 0.801 | 0.823 | 0.841 | 0.885 |
-| | Ultrasound Analysis | 0.723 | 0.745 | 0.762 | 0.741 |
-| | Cardiac Echo Analysis | 0.756 | 0.778 | 0.795 | 0.827 |
-| | Dental Radiograph | 0.734 | 0.756 | 0.773 | 0.755 |
+| **Segmentation Tasks** | Tumor Segmentation | 0.823 | 0.841 | 0.855 | 0.856 |
+| | Organ Detection | 0.891 | 0.902 | 0.911 | 0.909 |
+| | Brain MRI Segmentation | 0.812 | 0.825 | 0.838 | 0.836 |
+| **Classification Tasks** | Lesion Classification | 0.756 | 0.771 | 0.785 | 0.779 |
+| | Skin Lesion Classification | 0.834 | 0.848 | 0.862 | 0.856 |
+| | Pathology Classification | 0.789 | 0.802 | 0.815 | 0.809 |
+| | Chest X-ray Analysis | 0.867 | 0.879 | 0.891 | 0.886 |
+| **Detection Tasks** | Bone Fracture Detection | 0.723 | 0.738 | 0.752 | 0.749 |
+| | Mammography Detection | 0.801 | 0.815 | 0.828 | 0.829 |
+| | Dental X-ray Analysis | 0.778 | 0.792 | 0.805 | 0.799 |
+| | Cardiac Imaging | 0.845 | 0.858 | 0.871 | 0.869 |
+| **Advanced Analysis** | CT Scan Analysis | 0.856 | 0.869 | 0.882 | 0.879 |
+| | Ultrasound Segmentation | 0.734 | 0.748 | 0.761 | 0.756 |
+| | Retinal Screening | 0.812 | 0.826 | 0.839 | 0.836 |
+| | Spine Alignment | 0.767 | 0.781 | 0.794 | 0.789 |
 
 </div>
 
 ### Overall Performance Summary
-MedVisionNet demonstrates superior performance across all medical imaging benchmark categories, with particularly strong results in radiological and ophthalmological tasks.
+MedVisionNet demonstrates exceptional performance across all evaluated medical imaging benchmarks, with particularly strong results in segmentation and classification tasks.
 
 ## 3. Clinical Integration & API Platform
-We provide a clinical integration interface and API for healthcare institutions. Please contact our medical AI division for deployment options.
+We provide a HIPAA-compliant API for clinical integration. Please contact our enterprise team for deployment options.
 
 ## 4. How to Run Locally
 
-Please refer to our clinical documentation for information about running MedVisionNet in your environment.
+Please refer to our code repository for more information about running MedVisionNet locally.
 
-Model usage recommendations:
+Key usage recommendations for MedVisionNet:
 
-1. DICOM input is fully supported with automatic preprocessing.
-2. Multi-modality fusion can be enabled for comprehensive analysis.
+1. Image preprocessing with DICOM standardization is supported.
+2. Batch inference is optimized for throughput in clinical workflows.
 
-The model architecture of MedVisionNet-Lite is optimized for edge deployment while maintaining diagnostic accuracy.
+The model architecture of MedVisionNet-Lite is identical to its base model, optimized for edge deployment in medical devices.
 
-### Input Preprocessing
-We recommend the following preprocessing pipeline:
+### Input Specifications
+Images should be preprocessed to the following specifications:
 ```
-preprocess_config = {
-    "image_size": 512,
-    "normalize": "imagenet",
-    "augmentation": False
-}
+input_size = (512, 512)
+normalization = "medical_standard"  # Uses Hounsfield units for CT
+channels = 1  # grayscale for most modalities
 ```
 
 ### Inference Configuration
-We recommend setting the confidence threshold to 0.7 for clinical applications.
+We recommend the following configuration for optimal results:
+```python
+config = {
+    "confidence_threshold": 0.85,
+    "nms_threshold": 0.5,
+    "use_tta": True,  # Test-time augmentation
+    "ensemble_mode": "weighted_average"
+}
+```
 
-### DICOM Processing Template
-For DICOM file processing, use the following template:
+### DICOM Integration
+For DICOM file processing, use our provided utilities:
+```python
+from medvisionnet import DicomProcessor
+
+processor = DicomProcessor()
+image = processor.load_dicom("path/to/study.dcm")
+prediction = model.predict(image)
 ```
-dicom_template = \
-"""[study_id]: {study_id}
-[modality]: {modality}
-[body_part]: {body_part}
-[pixel_data_begin]
-{pixel_array}
-[pixel_data_end]
-{clinical_query}"""
+
+For multi-slice CT analysis, we recommend the following approach:
+```python
+ct_processor = DicomProcessor(modality="CT")
+volume = ct_processor.load_series("path/to/ct_series/")
+predictions = model.predict_volume(volume, slice_thickness=1.0)
 ```
 
 ## 5. License
-This code repository is licensed under the [Apache 2.0 License](LICENSE). MedVisionNet is intended for research and clinical decision support only.
+This code repository is licensed under the [Apache 2.0 License](LICENSE). The use of MedVisionNet models is also subject to the [Apache 2.0 License](LICENSE). The model is approved for research use; clinical deployment requires additional validation.
 
 ## 6. Contact
-For clinical inquiries, please contact medical@medvisionnet.ai.
+If you have any questions, please raise an issue on our GitHub repository or contact us at medical-ai@medvisionnet.ai.

config.json

@@ -1,4 +1,19 @@
 {
-    "model_type": "vit",
-    "architectures": ["ViTForImageClassification"]
-  }
+  "model_type": "vit",
+  "architectures": [
+    "ViTForImageClassification"
+  ],
+  "hidden_size": 768,
+  "num_hidden_layers": 12,
+  "num_attention_heads": 12,
+  "intermediate_size": 3072,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.0,
+  "attention_probs_dropout_prob": 0.0,
+  "initializer_range": 0.02,
+  "layer_norm_eps": 1e-12,
+  "image_size": 512,
+  "patch_size": 16,
+  "num_channels": 1,
+  "qkv_bias": true
+}

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:007d078aa561745802b0ecd4b1d1720922db71444bc4130f5830a0af69fc72de
-size 10240
+oid sha256:7cb65aecc1a4620d34d4069cc12aa569580dc1564151c3a97edd2060cc795b56
+size 39