Extract features from 3D medical or scientific images.
Volumetric image processing
Keras · Class
layers.Conv3D
Conv3D applies a learned kernel across three spatial or temporal dimensions to produce volumetric feature maps. It supports configurable filters, kernel size, strides, padding, dilation, activation, bias, grouping, initialization, regularization, and constraints.
How it works
Conv3D slides a learned kernel across three spatial or temporal dimensions and emits one value per output filter at each location. It can add a bias and apply an activation after convolution; the input and output are rank-5 tensors, with channel placement determined by data_format.
Syntax
keras.layers.Conv3D(filters, kernel_size, strides=(1, 1, 1), padding="valid", data_format=None, dilation_rate=(1, 1, 1), groups=1, activation=None, use_bias=True, kernel_initializer="glorot_uniform", bias_initializer="zeros", kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs)Parameters
| Name | Type | Description |
|---|---|---|
filters | int | Number of output filters. |
kernel_size | int or tuple/list of 3 integers | 3D convolution window size; an integer or three integers. |
strides | int or tuple/list of 3 integers | Convolution stride; an integer or three integers. |
padding | string | Padding mode: valid or same. |
data_format | string or None | Channel ordering, such as channels_last or channels_first. |
dilation_rate | int or tuple/list of 3 integers | Dilation rate; an integer or three integers. |
groups | int | Number of groups for grouped convolution. |
activation | activation function or None | Activation function. |
use_bias | bool | Whether to include a bias vector. |
kernel_initializer | initializer | Initializer for the kernel. |
bias_initializer | initializer | Initializer for the bias. |
kernel_regularizer | regularizer or None | Regularizer for the kernel. |
Returns
A 5D tensor representing activation(conv3d(inputs, kernel) + bias).
Key points
- The kernel operates across three spatial or temporal axes.
- filters sets how many output feature channels the layer produces.
- kernel_size accepts one integer or three integers describing the window.
- strides controls the step size along each convolution axis.
- dilation_rate spaces kernel elements for dilated convolution.
- use_bias adds a trainable bias vector when enabled.
- activation is applied after convolution and bias addition when provided.
Common mistakes
- Do not combine a stride greater than 1 with a dilation rate greater than 1.
- The documented padding choices are only valid and same.
Example: Basic 3D convolution
import keras
x = keras.Input(shape=(16, 32, 32, 3))
y = keras.layers.Conv3D(32, 3, activation="relu")(x)
model = keras.Model(x, y)
print(model.output_shape)Creates a 3D convolution with 32 output filters and a cubic kernel of size 3, then applies ReLU.
When to use it
Learn local patterns across depth, height, and width or time-related axes.
Video or spatiotemporal modeling
Build feature-extraction stages for volumetric classification or segmentation models.
3D encoder networks
Interview questions
QWhat does Conv3D compute?
Conv3D performs convolution over three dimensions and produces one output channel for each filter. With bias enabled, it computes the convolution, adds the bias, and then applies the optional activation.
Follow-up: What constraint applies to strides and dilation_rate? They cannot both use values greater than 1.
QWhat do filters and kernel_size control?
filters determines the output-space size, or number of convolution filters. kernel_size determines the 3D window dimensions and may be one integer or a sequence of three integers.
Follow-up: How can one integer be used for kernel_size? It specifies the same window size in all three dimensions.
QWhat is the output of Conv3D?
The layer returns a rank-5 tensor representing activation(conv3d(inputs, kernel) + bias). The exact spatial arrangement depends on padding, strides, data_format, and the input shape.
Follow-up: What changes when data_format is channels_last? The channels dimension is placed last in the rank-5 tensor.
Related
Conv1DConv2DConv3DTransposeConvLSTM3D