Synthetic Spermatozoa Video Sequences Generation Using Adversarial Imitation Learning

Abstract

Automated sperm sample analysis using computer vision techniques has gained increasing interest due to the tedious and time-consuming nature of manual evaluation. Deep learning models have been applied for sperm detection, tracking, motility analysis, and morphology recognition. However, the lack of labeled data hinders their adoption in laboratories. In this work, we propose a method to generate synthetic spermatozoa video sequences using Generative Adversarial Imitation Learning (GAIL). Our approach uses a parametric model based on Bezier splines to generate frames of a single spermatozoon. We evaluate our method against U-net and GAN-based approaches, and demonstrate its superior performance.

This research work has been supported by project TED2021-129162B-C22, funded by the Recovery and Resilience Facility program from the NextGenerationEU and the Spanish Research Agency (Agencia Estatal de Investigación); and PID2021-128362OB-I00, funded by the Spanish Plan for Scientific and Technical Research and Innovation of the Spanish Research Agency.

Appendix

In this section we detail the network architectures

GAIL.  We use a PPO agent as a generator and a CCN as a discriminator.

• PPO   is an actor-critic RL agent that comprises two neural networks referred to as the actor and the critic. Both networks shared the same backbone, consisting of two convolutional layers alternated with max pooling layers, followed by flatten and passed through a hidden dense layer with ReLU activation. The actor network estimates the policy $\pi _\theta$ so its output layer comprises five dense neurons with tanh activation. The critic network evaluates the actor’s decisions, so its output layer just has one neuron with linear activation.

• Discriminator   is similar to the critic but with a sigmoid activation at the output. To prevent overfitting we incorporate dropout after each convolutional and dense layer.

WGAN.  It has the conventional architecture.

• Generator   has an input layer of size $100\times 1$ followed by three transposed convolution layers. We alternate these layers with batch normalization and ReLU activation.

• Discriminator   has an input layer of size $28\times 28$. Then, it has the same architecture as GAIL’s discriminator but with three convolutional layers with a single dense neuron with linear activation at the output.

U-Net.  It receives an input tensor of size $5\times 28\times 28$ corresponding to the five preceding frames, $[s_{t-4},\ldots , s_t]$. Then, it compresses the input tensor using four convolutions to a size of $64\!\times 4\!\times 4$. The image reconstruction is made by three successive transposed convolutions, resulting in an output tensor of size $1\times \!28\!\times 28$, corresponding to the next frame $s_{t+1}$. All hidden activations are ReLU.

CWGAN.  We use the same U-Net architecture for the generator and a CNN-based discriminator.

• Generator   produces an image for $s_{t+1}$ conditioned on the input in $s_t$.

• Discriminator   has an input layer of $6\!\times 28\!\times 28$, corresponding to the five preceding frames concatenated with the one predicted by the generator. The architecture is the same than WGAN’s Discriminator.