#58 - Can AI find sperm that the human eye misses?

A pregnancy from two viable sperm recovered by AI sounds impossible until you walk through the workflow step by step. We start with the reality that male factor infertility can drive a huge share of infertility cases, and we talk about why “no sperm found” is not just a lab result but a years-long clinical and emotional grind. When the only path forward involves invasive sampling and hours of microscope time, even the best teams are fighting biology, fatigue, and the limits of manual search.

Then we unpack the STAR system, a sperm tracking and recovery approach that pairs computer vision with physical automation. We explain how modern object detection, including YOLO-style models, can scan microscopy imagery at massive scale, spotting sperm amid blood cells and tissue where a human could easily miss them. But the real leap is that it does not stop at detection. Microfluidic chips with hair-thin channels, gating mechanisms, and robotic handling help isolate and recover sperm quickly so they can be used in IVF workflows like intracytoplasmic sperm injection (ICSI).

Finally, we dig into the part that makes this feel real: the numbers and the clinical outcome. Millions of images scanned, a few sperm recovered, embryos created, and a positive pregnancy after a 19-year infertility history. We also wrestle with a key AI in medicine question: when a task is close to “sperm or no sperm” and the system is both fast and highly accurate, how much should we demand interpretability versus validation and results?

References:

First clinical pregnancy following AI-based microfluidic sperm detection and recovery in non-obstructive azoospermia
Suryawanshi et al.
The Lancet (2025)

STAR (Sperm Tracking and Recovery) System

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
 Licensed under Creative Commons: By Attribution 4.0
 https://creativecommons.org/licenses/by/4.0/

Transcript

Needle In Haystack Hook

SPEAKER_01 0:00

Could AI find the needle in the haystack? In this case, the needle was sperm, and the result was pregnancy.

Why Male Factor Infertility Is Hard

SPEAKER_00 0:16

Hello and welcome back to Coding Cure. My name is Vasant Sarathi. I'm a cognitive scientist and AI researcher, and I'm here with Laura Hagopian.

SPEAKER_01 0:24

I'm an emergency medicine physician. And today we're talking about the first clinical pregnancy that happened after an AI-based sperm detection method. It's pretty cool, actually, if you think about it. Incredible. Like amazing. Yeah.

SPEAKER_00 0:42

Yeah.

SPEAKER_01 0:43

A really good use case. And I think maybe we'll start off. I can start off by talking about, you know, male factor infertility being pretty common. Yeah.

SPEAKER_00 0:52

It has everything, right? Zorgon interrupt you, but I'm just like expressing my excitement here because it has a problem that people are dealing with. You know, how difficult it is to do right now. AI comes in and helps. But it's not just AI. There's like a physical manifestation, a robotic piece to it too, which is very cool, I think.

SPEAKER_01 1:08

Yeah, for sure. Exactly. Like they are using some sort of system to find the sperm and then like siphon them off to use them to fertilize the eggs, right? So it's like more than one step where AI and robotics are involved in this process. Um, and in a way that makes it like more efficient and easier, which is cool.

SPEAKER_00 1:30

Yeah, amazing.

SPEAKER_01 1:32

So male factor infertility can account for up to 40% of cases of infertility. Um, and oftentimes, you know, it's because there's not enough sperm or there's like no sperm that can be found in the semen. Um, and so this is hard to treat, right? It's like, what do you what do you do in this case? Um, oftentimes there's like years and years of trials, um, invasive procedures that may happen where they like take samples from the testicles themselves, right? Um, and it can be difficult when you're trying and trying and trying, and it's so it's so hard to get the sperm that you want to fertilize the egg. Right. Um, it can be really hard for couples. And so before this AI method and current sort of standard of care is there are a few different ways they go about this, but one common thing that happens is actually extracting the sperm from the testicle itself.

SPEAKER_00 2:32

Wow. That's seems complicated.

SPEAKER_01 2:36

Yeah, I thought you were just gonna say, like, ouch.

SPEAKER_00 2:38

Well, that too, but um But the thing is But it's also the scale of this, right? The sperm, I mean, like you're getting it out and it's like tiny, tiny, tiny, and you need microscopes to see them. And it seems like, and you can only look at so big a sample in each microscope setting.

SPEAKER_01 2:56

Yeah, I mean, you're kind of like so that that's sort of the next step. So the first step is like, oh, there's not there's not enough, or there's almost no sperm in the semen sample. Then you go to the testicle and you're like, hey, is there is there anything here? Is there more here? Um,

The Manual Microscopy Bottleneck

SPEAKER_01 3:10

because it maybe it's present in the testicles in places that it wasn't coming out, or not enough of it was coming out. And so that's what they'll do is they'll they'll sort of like dissect away, or they can take a needle and take a sample, or they take multiple samples. Um, and they do it with like a, yeah, like a high power operating microscope to kind of look at the tubes that are most likely to contain sperm inside the testicles and pull those out. But then the next step, and this is where I think the AI starts to come in, is like, hey, now that you have these samples, you want to look for the actual sperm in them under the microscope. And that is like a very labor-intensive, hands-on, painstaking process, which is like, great, can we automate this? Right. Right. When when you think about AI coming in, it's like someone is looking under the microscope and scanning and scanning and scanning for hours and hours this sample, which there's other stuff there other than the sperm, right? And they're looking to see, you know, okay, there's blood cells, there's other tissue. Can we find a sperm that's like moving that we think is gonna work? Um, and because the sample is large and you're looking at it blown up under the microscope, this is this is like labor a laborious process. Right. Um, and so then what they'll do is if they do find some sperm, in many cases they they they may not, but if they do find it, they'll try to get it out and then they'll um use it to fertilize an egg.

SPEAKER_00 4:37

Okay.

SPEAKER_01 4:38

And normally sperm, like we let it just fertilize an egg on its own, like it swims inside of it. But in this case, because you're gonna have so few sperms, you're actually gonna inject it into the egg cell. Like, but you can't. Like you're gonna like guarantee that it gets in there.

SPEAKER_00 4:53

Unfathomable, I shouldn't say unfathomable, but it's hard for me to even imagine how they pull out one in the sample. Like it feels it's microscopic, right? I mean, they don't have microscopic tweezers or something. I don't understand, I don't, I'm trying to figure out how this works.

SPEAKER_01 5:06

Yeah. Wow. This is absolutely not my specialty, so I can't I can't necessarily tell you the details of that one, but when they do pull it out, what they'll do is they'll um they'll actually like inject it into the egg because because of the small number that of sperm that are available, you you're not guaranteed that the sperm's even gonna get into the egg. Right.

SPEAKER_00 5:28

So you want to just like reduce any chance that it misses or whatever.

SPEAKER_01 5:32

Exactly. So what I think is cool about this paper is they were like, well, and in this technique in general, this sperm tracking and recovery system, the star system, which was developed where out of Columbia, right? Is that they were like, well, this long laborious process of looking under the microscope to try to find, you know, this the sperms that we think could work, you know, where are the functional sperm that we can use? Right. Instead of have taking someone looking under the microscope for hours and hours, potentially, you know, missing some human error. And also, like, can you imagine doing that for eight hours?

SPEAKER_00 6:11

Wow. Yeah. Can we that's what that's what's happening right now? People are looking at it for eight hours or whatever, right?

SPEAKER_01 6:16

Yeah. So then the question is like, can we automate a process like that or can we use AI to kind of augment that? So walk me through kind of how this star system works.

SPEAKER_00 6:28

Well, I can begin with the part that is the AI piece, which is um we've had image detectors for a while now. 2012, 2010, around is around the time when um the first image detectors started really coming into play. You can you had um what's called convolutional neural networks that in theory have been around since the 90s, but uh they were made possible and sort of scaled up because we have GPUs in the early 2000s and the mid-2000s, and by 2012, people had applied those techniques to for images. And the convolutional neural networks are a little different from just like normal neural networks. Uh they have a specific function that they can do, which is be able to um I identify when things are close to one another. And that's great for images. Like you know in a car, if you s if you take a picture of a car, if you are uh by the headlight, you know, like and you and you kind of walk from left to right. Imagine the car is kind of sideways, and you're by the headlight, and as you move, you know that the wheel is coming up next. There's like space spatial uh relationships that exist in a visual image. Yeah, okay, that means so convolutional neural networks exploit that spatial structure and allow you to um identify objects and things like that in the world from the images. And so they've been around for a really long time and they've been perfected for a really long time, even after the LLMs and Transformers and all those other architectures came out. Um convolutional neural nets are still very good and for many reasons. And one of the most popular ones is something called YOLO, or you look only once. Not you only live once. Right. You only look once, right? Now you only live once. It's it's a one-pass, one-shot. What that means is sort of the in in the older prior to YOLO days, um, these detectors ran in sort of two stages. There was a first stage where they proposed kind of regions that might look like they contain objects. And then once they did that, they had a separate stage of actually classifying, going through each of the regions and then figuring out what the objects are.

SPEAKER_01 8:26

So in this case, they might must not be doing that because they're only looking one time, not two times.

Computer Vision Meets Microfluidics

SPEAKER_01 8:30

Yeah.

SPEAKER_00 8:30

So they reframed the whole detection process as a single problem. And the neural network looks at the entire image once and simultaneously predicts all the objects and labels. Um, and it's a trick that makes the whole thing so much faster that they're actually really good for real-time video. So they can detect it between 30 to 100 frames per second.

SPEAKER_01 8:50

Oh my gosh.

SPEAKER_00 8:50

So that's faster than video. So that means every video frame is basically they you run it through the YOLO detector and then you get an updated box, right? So you can basically track this thing moving around, right? Because you're going with the video.

SPEAKER_01 9:02

Um, and so in this case, they're like, hey, we want to figure out where the sperm is. Yeah, yeah. And that can do it. I mean, I'm looking at the paper, it's like the rate of 1.1 million images per hour. So they like subdivide, my understanding is they like like subdivide it into like frames and then try to detect where inside those frames, like almost like a grid where the sperm are.

SPEAKER_00 9:29

Yes, yes. And that frame uh framing process was not necessarily for the detection of YOLO because we know that it looks at the whole image, but it was more so for the next step, which is once you've found the sperm, how do you get it out? And that was pretty cool too. They had, you know, a way to like channel the sperm out using the microfluidics uh basis under you know, kind of in where the sperm is. So that was pretty cool too.

SPEAKER_01 9:54

So instead of like having a human involved in any of this, like it's doing both of those steps. It's it's very quickly looking at the whole image, the entire sample, and figuring out where are the sperm, which which objects are sperm here. And then when it's finding a sperm, it's able to like use a gating mechanism to sort it out, and then they can then transfer it to the lab for someone to do that intracytoplasmic injection where they actually inject the sperm into the egg.

SPEAKER_00 10:25

Yeah, no, and what's cool about this is that that chip that the sperm's on, or the other, or the sample is on, um, is engraved with channels that are as thin as a human hair. They're really, really small, and that allows them to isolate portions of the of the sample uh that contain certain sperm sperm cells. And like you said, the robotic arm then removes the individual sperm cells. And what I think is really cool is that it does this within milliseconds.

SPEAKER_01 10:48

That's crazy. Um I mean, you think about this job that takes hours and hours and hours, and someone just like looking under the microscope, and instead you're like, okay, well, now it only takes one hour for the AI to do this.

SPEAKER_00 11:03

Yes, and critically, the AI is super accurate. So like these AI systems have 95 plus percent accuracy in uh detecting uh detecting sperm. And um they have very high uh you know accuracy, and also because they're fast, you know, they're this essentially what this does is then avoids all the other processes that people used to do before, which was like centrifugation, um lasers and dyes and all kinds of other things that could maybe damage the sperm. So that you know, this takes out all of that and preserves the viability of the sperm uh in the process. So I think there's so many cool innovations here. I mean, I feel like we you know we can talk about all of them separately in a separate podcast, but there's so much here. Um and the speed is is is speed and accuracy are just incredible. Um and it can basically they said it can generate approximately 40 times more um, you know, it can detect 40 times more sperm than compared to the human trained human lab techs, which is incredible.

SPEAKER_01 12:02

That is incredible. And I feel like this is an awesome use case because you because you have this highly manual process that, you know, if you're staring at a sample for eight hours, of course you're gonna miss some stuff, right? I would for sure. Um, and you're like, well, how can we, how can we automate this or how can we use AI to augment what we're doing? And not that humans aren't involved at all in the process, right? Like they're gonna be involved in the intracytoplasmic injection or whatever else.

SPEAKER_00 12:27

Yeah.

SPEAKER_01 12:27

But this piece of it is so promising. And I think what's really cool, and we'll link this uh paper in the show notes, is that they actually have a they have a case of like, hey, this resulted in a pregnancy, right? It's not just like theoretical. We didn't just like find the sperm. Um, we did more than that, and it it has resulted in a positive pregnancy.

SPEAKER_00 12:48

And I think that example is mind-blowing if you look at the numbers. So is uh from a three and a half milliliter um semen sample, the star system scanned two and a half million images in approximately two hours and identified two viable sperm cells. And those were used to create two embryos.

SPEAKER_01 13:07

That's crazy, right? You have this huge sample, tons of images and two sperm that were mobile. But I think they found uh five non-motile sperm and two motile sperm. So they used the two that were motile, they did the intracytoplasmic injection into the mature eggs, and uh then they transferred the embryos.

SPEAKER_00 13:30

Yeah. And and and this is, you know, you might think, okay, how fast do humans do this? I think that there was also another documented uh case uh in which they had a this is not a clinical case, a pre-clinical clinical case, but the uh the skilled technician searched a sample for two days and found no sperm. And the star system analyzed the same sample in one hour and identified 44 sperm cells, which again, uh another piece like you know, this is the we talked about how it can find things, but it can also find things faster and better than a lap tech can. Um and you know, this is not one of those cases where um to me at least, you know, from our AI perspective and from our medical, like our podcast perspective, we've talked a lot about explainability, interpretability, AI is making these reasoning choices and coming to conclusions. This is less of that, right? This is more like, is it there or not? Yeah. And if it's there, does it work or not?

SPEAKER_01 14:23

Yeah, exactly. It's like it's like binary, sperm or no sperm. Or there's a second level, like, is the sperm motile or not? But yeah.

Real-World Pregnancy And The Data

SPEAKER_01 14:30

And I think this is particularly interesting because the couple that this that that had this successful pregnancy, they had a 19-year history of infertility. They had not been able to get pregnant for 19 years. They had had all this testing done, all of uh, you know, this huge urology evaluation. Um, they'd done all of these manual sperm searches, they tried to like extract the sperm, you know, from the testicles, and nothing had worked, right? They'd been to like tons, like more, like a few different fertility centers. Um, they had tried to do different methods of fertilization, they tried to do IVF, etc., and nothing had resulted in a pregnancy. And here, I mean, you can you can see like he didn't have many sperm at all, but they were able to recover some from a plain old semen sample, too. They they didn't have to take it from the testicle, right? They were able to find the the two viable sperm, fertilize the eggs, get them into the female partner, and have a positive pregnancy test. And they had an ultrasound that showed, hey, the heart, the heart is beating.

SPEAKER_00 15:44

Right. It's so cool. Yeah, yeah. That that is incredible. And again, I want to go back to the question of trust and and explainability and interpretability in these systems and why we don't care as much about that in this particular case. It is interesting if you think about it, right? Why why don't we care how explainable was the system in terms of figuring out if the sperm, I mean, it's a scientific curiosity, sure. How did it figure out sperm versus no sperm? But uh beyond that, from a use case perspective, we're less interested. Is it because the accuracy is so high, or is it because it resulted in a real world benefit and that or that the benefits outweigh the cost of getting it wrong? I mean, the the what would happen if you get this wrong, right? If if it identified uh a bad sperm or it couldn't find one, right? It feels like the cost of that is much lower because it was so much harder for humans to begin with, right? To do it. And humans are not doing it very well to begin with, right? So uh to me, what's interesting is why do we care less about we I I don't care as much about its interpretability. And I don't know if you feel the same way.

SPEAKER_01 16:47

Yeah, I mean, I feel like the the ultimate goal in this case is to isolate sperm and have a pregnancy,

Trust, Accuracy, And When To Move Fast

SPEAKER_01 16:55

right? And people are coming in with infertility issues, male factor infertility issues, problems where they don't have enough sperm or any sperm that people can find, right? And then there's a goal they're trying to get to. So then it's like, well, how do you get from A to B?

SPEAKER_00 17:11

Right.

SPEAKER_01 17:12

And there are lots of different paths, right? And maybe they've tried 17 paths already, and this was the 18th. And it doesn't matter how you get from A to B. You just need to get there. And they hadn't gotten there in the first 17. And the 18th got them there. And I'm sure they don't care about, hey, how interpretable is this? Because they achieved their goal. Right. And that part I think is cool. But then you back up and you're like, well, this is a highly manual process. This piece of it could be automated. Maybe there's ways to like injure the sperm less once they're found, etc. It's like, I'm not sure we need a fully explainable or an interpretable system here.

SPEAKER_00 17:51

Right. I I think I was coming at it more from the angle of trust. Like we trusted, would you trust the system? And the answer seems to be, I mean, at least for me, it seems to be yes. If it works, yeah. I think there's no cost, there's no real cost of it getting it wrong, right? I think that's kind of the point. Unless that there is a cost of it getting wrong, the trust question really becomes important.

SPEAKER_01 18:09

Well, I suppose like you could, in theory, like say it didn't work well, say uh humans did a better job than this system. Then you could have different issues.

SPEAKER_00 18:19

And then you have a trade-off, right? Right. I can do this faster, but you know, with all these errors, or I can do it, you know, uh slower, but but more accurately. And so you don't have that trade-off here. It's accurate and fast.

SPEAKER_01 18:30

It's both. It's both of those things. So it feels like a like a no-brainer situation of a really great application of AI that could potentially help a lot of people.

SPEAKER_00 18:39

Yeah, exactly. Exactly.

SPEAKER_01 18:42

So very, very positive episode, really cool outcome and not not just theoretical, right?

SPEAKER_00 18:49

Yeah.

SPEAKER_01 18:50

Actually practical for people who have male factor infertility, which can can be 40% of infertility cases.

SPEAKER_00 18:57

Yeah, that's amazing.

SPEAKER_01 18:59

Yeah. All right. Well, we will see you next time on Coding Cure. Thank you for joining us.