Initial commit

.gitattributes

@@ -1,6 +1,7 @@
 *.7z filter=lfs diff=lfs merge=lfs -text
 *.arrow filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
 *.bz2 filter=lfs diff=lfs merge=lfs -text
 *.ftz filter=lfs diff=lfs merge=lfs -text
 *.gz filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,5 +1,5 @@
 ---
-title: Bone Age Greulich And Pyle
+title: Deep Learning Model for Pediatric Bone Age
 emoji: 💻
 colorFrom: red
 colorTo: blue

app.py

@@ -0,0 +1,91 @@
+import gradio as gr
+import timm
+import torch 
+import torch.nn as nn
+
+
+def change_num_input_channels(model, in_channels=1):
+    """
+    Assumes number of input channels in model is 3.
+    """
+    for i, m in enumerate(model.modules()):
+      if isinstance(m, (nn.Conv2d,nn.Conv3d)) and m.in_channels == 3:
+        m.in_channels = in_channels
+        # First, sum across channels
+        W = m.weight.sum(1, keepdim=True)
+        # Then, divide by number of channels
+        W = W / in_channels
+        # Then, repeat by number of channels
+        size = [1] * W.ndim
+        size[1] = in_channels
+        W = W.repeat(size)
+        m.weight = nn.Parameter(W)
+        break
+    return model
+
+
+class Net2D(nn.Module):
+
+    def __init__(self, weights):
+        super().__init__()
+        self.backbone = timm.create_model("tf_efficientnetv2_s", pretrained=False, global_pool="", num_classes=0)
+        self.backbone = change_num_input_channels(self.backbone, 2)
+        self.pool_layer = nn.AdaptiveAvgPool2d(1)
+        self.dropout = nn.Dropout(0.2)
+        self.classifier = nn.Linear(1280, 1) 
+        self.load_state_dict(weights)
+
+    def forward(self, x):
+        x = self.backbone(x)
+        x = self.pool_layer(x).view(x.size(0), -1)
+        x = self.dropout(x)
+        x = self.classifier(x)
+        return x[:, 0] if x.size(1) == 1 else x
+
+
+class Ensemble(nn.Module):
+
+    def __init__(self, model_list):
+        super().__init__()
+        self.model_list = nn.ModuleList(model_list)
+    
+    def forward(self, x):
+        return torch.stack([model(x) for model in self.model_list]).mean(0)
+
+
+checkpoints = ["fold0.ckpt", "fold1.ckpt", "fold2.ckpt"]
+weights = [torch.load(ckpt)["state_dict"] for ckpt in checkpoints]
+weights = [{k.replace("model.", "") : v for k, v in wt.items()} for wt in weights]
+models = [Net2D(wt) for wt in weights] 
+ensemble = Ensemble(models).eval()
+
+def predict_bone_age(Radiograph, Sex):
+    img = torch.from_numpy(Radiograph)
+    img = img.unsqueeze(0).unsqueeze(0)
+    img = img / img.max()
+    img = img - 0.5
+    img = img * 2.0
+    if Sex == 1:
+        img = torch.cat([img, torch.zeros_like(img) + 1], dim=1)
+    else:
+        img = torch.cat([img, torch.zeros_like(img) - 1], dim=1)
+    with torch.no_grad():
+        bone_age = ensemble(img.float())[0].item()
+    return f"Estimated Bone Age: {int(bone_age)} years, {int(bone_age % int(bone_age) * 12)} months"
+
+
+image = gr.Image(shape=(512, 512), image_mode="L")
+sex = gr.Radio(["Male", "Female"], type="index")
+label = gr.Label(show_label=True, label="Result")
+
+demo = gr.Interface(
+    fn=predict_bone_age,
+    inputs=[image, sex],
+    outputs=label,
+    )
+
+
+if __name__ == "__main__":
+    demo.launch()
+
+

fold0.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2db6d3fb26a05b916341574c83683017e4a04a1c0df8fda4a97ad2314b33f109
+size 81642981

fold1.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c806c2ccd21cb4f1d1102e86d8716ed67583f561d4eea6a1761ac4f9bf6a60b
+size 81642981

fold2.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cabdc105bb4c3239d1a57ceaaca4306096a017763c1ec1d23adacf6d8c0713ab
+size 81642981

requirements.txt

@@ -0,0 +1,5 @@
+numpy
+timm
+torch
+https://gradio-main-build.s3.amazonaws.com/e30af8813c3d76329cf4869fa87a902b2075c8cd/gradio-3.8.2-py3-none-any.whl
+