TensorFlow 2.2.0中CNN医学图像分割模型实现问题求助

Hey there! Let's break down each of your three problems and walk through practical fixes to get your model aligned with the literature and training smoothly.

1. Fixing Conv6 Output Size Mismatch (22x22 vs. 44x44) #

The root cause here is almost certainly inconsistent padding, pooling strides, or missing upsampling steps compared to the paper's architecture. Here's how to troubleshoot and correct it:

• Check padding settings: If you’re using padding='valid' in convolutional layers, this shrinks feature map sizes with each pass. Switch to padding='same' for layers where the paper specifies no spatial reduction—this keeps input and output dimensions identical when stride=1.
• Calculate dimension changes manually: Use the standard convolution output formula to verify each layer’s output size:

  output_size = (input_size - kernel_size + 2*padding) // strides + 1
  

  For example, if your conv5 outputs 22x22, add a 2x2 transposed convolution to upsample it to 44x44, matching the paper:

  conv6 = Conv2DTranspose(1, (2,2), strides=(2,2), padding='same')(conv5)
  

• Verify pooling layers: If the paper uses max pooling with strides=1 (instead of the default strides=pool_size), adjust your MaxPooling2D layers to avoid unnecessary spatial shrinkage.

2. Correctly Using PReLU Activation in Keras #

Unlike built-in activations like relu, PReLU is a trainable layer—not a string identifier you can pass directly to Conv2D. Here’s the right way to implement it:

from tensorflow.keras.layers import Conv2D, PReLU

# Correct implementation: Add PReLU as a separate layer after convolution
x = Conv2D(64, (3,3), padding='same')(input_layer)
x = PReLU()(x)  # Initializes a trainable PReLU layer with learnable alpha parameters

• Optional: If you want to share the learnable alpha parameter across spatial dimensions (height/width), use PReLU(shared_axes=[1,2])—this reduces trainable parameters, which can help with small medical image datasets.

3. Resolving Training Dimension Mismatch Error #

The ValueError happens because your model’s output (22x22x1) doesn’t match your training mask shape (353x353x1). Here’s how to fix it:

• Align model output with input/mask size: Medical image segmentation requires the output mask to match the input image size. Fix the first issue first by adding upsampling layers (like Conv2DTranspose or UpSampling2D) to scale feature maps back to 353x353.
• Double-check data shapes: Confirm both X_train and Y_train have the shape (num_samples, 353, 353, 1). Verify this with:

  print(X_train.shape, Y_train.shape)
  

• Remove unnecessary downsampling: If the paper’s architecture doesn’t shrink the input to 22x22, make sure you’re not adding extra pooling layers or using overly aggressive strides that reduce feature maps too much.

内容的提问来源于stack exchange，提问作者Artur Santos Nascimento