In vitro fertilisation procedure assisted with computer vision models for organic Senegalese sole (Solea senegalensis) culture

This study demonstrates that combining computer vision-based sperm selection with automated behavioral analysis to predict ovulation enables successful, hormone-free in vitro fertilization in organic Senegalese sole aquaculture.

The Gist

Imagine you are trying to bake the perfect cake, but you have a very strict rule: you cannot use any chemical leavening agents (like baking powder). You have to rely entirely on the natural rise of the dough.

Now, imagine that the "dough" is a fish egg, and the "baking powder" is a hormone injection usually given to fish to make them release their eggs. In the world of organic aquaculture (fish farming that follows strict natural rules), these hormone injections are forbidden.

The problem? Without the "chemical help," the fish often forget to release their eggs, or they release them at the wrong time. It's like waiting for a bus that might never come, or might arrive at 3 AM when you're asleep. If you miss the window, the eggs are useless.

This paper is about solving that "waiting game" for Senegalese sole, a type of flatfish that is delicious but hard to farm organically. The researchers came up with a two-part solution: The "Super-Sperm" Selection and The "Fish Whisperer" AI.

Part 1: The "Super-Sperm" Selection (Finding the Best Dad)

In the wild, fish just swim around and hope for the best. In a farm, the researchers realized that if they want to succeed without hormones, they need to be picky about the dads.

They used a high-tech microscope (called CASA) to look at the sperm of many male fish. Think of this like a talent show audition.

• Some sperm were "slow walkers" (low quality).
• Some were "average joggers."
• A few were "Olympic sprinters" (high quality).

They only picked the "Olympic sprinters" to try and fertilize the eggs. This ensured that if an egg did appear, it had the best possible chance of becoming a baby fish.

Part 2: The "Fish Whisperer" AI (Predicting the Bus Arrival)

Even with the best sperm, you still have a problem: When will the mom fish release her eggs?
In nature, the fish release eggs at night. But they don't send a text message saying, "Hey, eggs coming in 10 minutes!"

The researchers noticed that before the eggs come out, the fish act a little weird. They do two specific things:

1. Rest the Head (RTH): One fish rests its head on another fish's back. It's like a fish saying, "I'm ready to party."
2. Locomotor Activity (LA): The fish start swimming around frantically, like they are doing a pre-party dance.

The Old Way: A human farmer would have to stare at the fish tanks all night, every night, hoping to catch a glimpse of this behavior. It's exhausting and easy to miss.

The New Way (The AI):
The researchers installed underwater cameras and taught a computer (an Artificial Intelligence) to watch the fish 24/7.

• The AI learned to spot the "Head Rest" and the "Dance."
• It then used a mathematical formula (Logistic Regression) to calculate the probability that eggs would be released that night.

Think of the AI as a super-accurate weather forecast. Instead of saying "It might rain," it says, "There is an 85% chance of eggs tonight, specifically between 6 PM and 7 PM."

The Results: Did it Work?

The team tested this on two groups of fish.

• Without the AI: They just guessed when to collect eggs. They missed the boat most of the time.
• With the AI: The system predicted the ovulation nights with 82–85% accuracy.

Because they knew when to look, they were able to collect the eggs right when they were fresh. They successfully created baby fish without using a single hormone.

• Fertilization Rate: Up to 44% (which is great for organic farming!).
• Hatching Rate: 18% of those eggs grew into healthy baby fish.

The Big Picture

This study is like building a smart traffic light system for fish farmers.

• Before: Farmers were driving blind, hoping to catch the fish at the right moment.
• Now: The AI acts as the traffic light, turning green exactly when the "eggs" are ready to cross.

Why does this matter?

1. Organic Certification: It proves you can farm high-quality fish without using hormones, which is a huge deal for health-conscious consumers and organic labels.
2. Less Stress: The fish aren't being poked and prodded with needles; the AI just watches them from a camera.
3. Efficiency: Farmers don't waste time and money trying to fertilize eggs that aren't there yet.

In short, the researchers combined top-tier sperm selection with a computer that watches fish dance to solve a decades-old problem. They turned a game of chance into a reliable, natural science.

Technical Summary

1. Problem Statement

The aquaculture of Senegalese sole (Solea senegalensis) faces significant reproductive challenges in captivity, primarily the absence of natural courtship behaviors leading to reproductive failure.

• Current Limitation: Successful in vitro fertilization (IVF) currently relies on hormone-induced ovulation (using GnRHa agonists) to synchronize egg release.
• Regulatory/Ethical Constraint: The use of hormones is prohibited in organic aquaculture, creating a critical gap for sustainable production.
• Operational Challenge: Without hormones, ovulation timing is unpredictable. Ovulated eggs remain viable for only ~3 hours. Manual monitoring is inefficient, labor-intensive, and often misses this narrow window, leading to low fertilization success rates.

2. Methodology

The study employed a dual approach: validating a hormone-free IVF protocol and developing an AI-driven system to predict ovulation timing based on fish behavior.

A. Experimental Setup

• Subjects: Two experimental groups (Mix1 and Mix2) housed in 10 m³ tanks with recirculating systems at IRTA La Ràpita, Spain. Groups contained a mix of wild and cultured fish.
• Conditions: Natural photoperiod and temperature (manipulated weekly to stimulate spawning). No hormonal treatments were administered.
• Gamete Collection:
  • Sperm: Males were screened using Computer-Assisted Sperm Analysis (CASA) to select high-quality sperm (motility, velocity). Selected sperm was stored at 4°C.
  • Eggs: Females were checked for ovulation via gentle abdominal pressure. Eggs were collected for IVF trials.

B. Computer Vision & AI Pipeline

The core innovation was an automated system to detect reproductive behaviors and predict ovulation.

1. Data Collection: Underwater cameras recorded fish activity (19:00–20:00 h) during the spawning season.
2. Behavioral Definitions:
   • Rest the Head (RTH): One fish resting its head on another (courtship indicator).
   • Locomotor Activity (LA): Movement exceeding a speed threshold (4 pixels/frame).
3. Model Architecture:
   • Step 1 (Detection): A Convolutional Neural Network (CNN) was trained on a dataset of 5,000 augmented images (1,000 RTH, 1,000 Fish Alone/FA, plus augmentations) to detect fish and classify behaviors (RTH, FA, or "no fish").
   • Step 2 (Tracking): A tracking module monitored individual fish over time to quantify LA instances.
   • Step 3 (Prediction): A Logistic Regression Model used behavioral counts (RTH, LA) and time of occurrence as features to predict the probability of ovulation for the upcoming night.
   • Equation: $P(Ovulation) = \frac{1}{1 + e^{-(\beta_0 + \beta_1 RTH + \beta_2 LA + \beta_3 Time)}}$

C. IVF Protocol

• Fertilization: Performed using naturally ovulated eggs and selected high-quality sperm.
• Methods: Two methods were tested:
  1. Beaker Fertilization: Individual male-female pairings (4 million motile sperm per beaker).
  2. Massive Fertilization: Pooled sperm from multiple males.
• Evaluation: Fertilization rates and hatching success were monitored.

3. Key Results

A. Hormone-Free IVF Feasibility

• Success: The study successfully produced viable larvae without hormonal intervention.
• Performance:
  • Maximum fertilization rate: 44% (in massive fertilization trials).
  • Maximum hatching rate: 18%.
  • Key Finding: Fertilization success was highly dependent on sperm quality (selected via CASA) and timing. When eggs were collected outside the 3-hour viability window or from non-ovulated females, success dropped to near zero.

B. Behavioral Analysis

• Temporal Patterns: Both RTH and LA behaviors peaked between 18:00 and 19:00 hours.
• Correlation:
  • LA (Locomotor Activity): Showed a strong, significant positive correlation with ovulation nights. Ovulation nights had significantly higher movement counts compared to non-ovulation nights.
  • RTH (Rest the Head): Also elevated during ovulation nights but showed a weaker, non-significant negative association in the predictive model compared to LA.

C. AI Model Performance

• Combined Dataset: The logistic regression model achieved an 85% accuracy in predicting ovulation nights.
  • AUC (Area Under Curve): 0.95.
  • Recall: 100% for non-ovulation nights; 77% for ovulation nights.
• Group Specifics:
  • Mix1: 75% accuracy (Optimal threshold: 0.5).
  • Mix2: 82% accuracy (Optimal threshold: 0.28).
• Significance: The model successfully identified ovulation events several hours in advance, allowing for targeted gamete collection.

4. Key Contributions

1. Proof of Concept for Organic Aquaculture: Demonstrated that Senegalese sole can be bred via IVF without hormones, achieving commercially viable fertilization and hatching rates, provided sperm quality is optimized.
2. AI-Driven Ovulation Prediction: Developed a non-invasive, automated system using CNNs and logistic regression to predict ovulation with >80% accuracy, solving the "timing" bottleneck of hormone-free breeding.
3. Behavioral Biomarkers: Validated that Locomotor Activity (LA) is a superior predictor of imminent ovulation compared to static courtship behaviors like RTH in this species.
4. Operational Framework: Proposed a workflow where AI monitoring guides manual intervention, transforming random sampling into a managed, efficient protocol.

5. Significance and Impact

• Sustainability: This approach enables the production of organic Senegalese sole, complying with strict EU regulations prohibiting hormonal treatments.
• Economic Viability: By predicting ovulation, aquaculture facilities can reduce wasted resources (labor, sperm, eggs) associated with random checking. The model shifts the process from reactive (checking after the fact) to proactive (preparing for the event).
• Animal Welfare: The system is non-invasive, eliminating the need for frequent handling or stress-inducing hormone injections.
• Scalability: The integration of deep learning (CNN) for behavior detection and statistical modeling (Logistic Regression) offers a scalable template for reproductive management in other fish species where natural spawning is difficult.

Conclusion: The study concludes that while hormone-free IVF is technically feasible, its commercial success relies entirely on the ability to predict the narrow ovulation window. The integration of computer vision and machine learning provides the necessary predictive capability to make organic flatfish aquaculture a practical reality.