Upload folder using huggingface_hub

README.md

@@ -0,0 +1,50 @@
+---
+license: other
+library_name: tensorflow
+tags:
+- image-segmentation
+- medical-imaging
+- dental
+- unet
+---
+
+# Dental Tooth Segmentation EfficientNet U-Net
+
+This repository contains the trained Keras model for the course assignment
+"Development of a Dental Tooth Segmentation System Using U-Net and Flask".
+
+The checkpoint is an EfficientNetB0 encoder U-Net trained and fine-tuned on
+dental panoramic X-ray segmentation data. It predicts a binary tooth-region mask
+from a panoramic dental radiograph.
+
+## Model File
+
+- `best_model.keras`
+
+## Inference Recipe
+
+- Resize image to 256x512.
+- Convert grayscale radiograph to three channels for the EfficientNet encoder.
+- Predict with horizontal-flip test-time augmentation.
+- Use threshold 0.65 for the final combined test result.
+- Remove connected components smaller than 32 pixels.
+
+## Metrics
+
+Final combined held-out test result:
+
+- Precision: 89.54%
+- Recall: 91.93%
+- F1/Dice: 90.72%
+- IoU: 83.02%
+- Pixel accuracy: 96.99%
+
+HITL held-out split:
+
+- Precision: 89.45%
+- Recall: 91.41%
+- F1/Dice: 90.42%
+- IoU: 82.51%
+
+This model is for education and demonstration only. It is not a medical device
+and must not be used for clinical diagnosis.

metrics.json

@@ -0,0 +1,34 @@
+{
+  "model": "combined_teeth_efficientnetb0_finetuned",
+  "model_file": "best_model.keras",
+  "image_size": {
+    "height": 256,
+    "width": 512
+  },
+  "inference": {
+    "threshold": 0.65,
+    "tta": "horizontal_flip",
+    "min_component_size": 32,
+    "closing_iterations": 0,
+    "opening_iterations": 0,
+    "fill_holes": false
+  },
+  "combined_held_out_test": {
+    "sample_count": 164,
+    "precision": 0.8953982106231072,
+    "recall": 0.9193066372231264,
+    "f1": 0.9071949292034751,
+    "dice": 0.9071949292034751,
+    "iou": 0.8301525619223543,
+    "accuracy": 0.9699334400456127
+  },
+  "hitl_held_out_test": {
+    "sample_count": 90,
+    "precision": 0.8944768761200106,
+    "recall": 0.9140709919470774,
+    "f1": 0.9041677910384308,
+    "dice": 0.9041677910384308,
+    "iou": 0.8250969296615555,
+    "accuracy": 0.9690822177463108
+  }
+}