Should AI Replace Radiology Second Readers? + Webcam HR Tracker

keep testing this tool and that's the the one that does your heart rate from the camera. You need to kind of remove your hair if you have any and so your board is is visible. Alternatively, you can do a manual ROI box, which actually might work uh better or worse. I don't know. You can make it a bit bigger. If it's a bit bigger, it will work a bit better, but then you have to make sure you're not moving. And we'll have a measurement. Is that my heart rate? It's a bit elevated. 72. Um, not sure if the G is correct. If G wants to support my videos, they're more than welcome. As soon as I start talking, as soon as a human, well, supposedly I'm human, you wouldn't know, would you? starts talking the engagement drops. So that's interesting. Wait, let's turn on the automated arrow for a sec. Still have the audio going. Do I have enough natural light? Okay, I have a reading. Just have to stop moving. Yeah, the signal quality is patchy. Probably don't have enough light in the room or something. Yeah, there's not enough light coming from the outside. I don't know if this this will help or not. The hairline in the thing doesn't help. Yeah, it's not a great uh reading. Somehow we get a bit of noise. Wait, any more natural light? Okay, let's check it again. Don't get a strong signal. Yeah, I'm not sure. Not getting a strong signal today. Okay, let's check it now. Okay, now we're talking. Well, I can see it. Get 75. What's me moving? But yeah, there's a strong signal. Let's turn on the signification cuz also the light is going straight into my eyes. is 77. Yeah, my G shows 78. So that's actually accurate. Accurate enough. So now I'm not taking the measurement for this from my smartwatch. Taking it uh directly from just the web camera alone. And it's accurate within the one one to plus - one bits per minute. This light is killing me in my face. Should try it again with the light moved behind my screen. By the way, if you have something flashing on the screen, it can mess up the recording as well. Yeah, the signal to noise is worse. So, there's less signal lower signal to noise ratio is lower. Let's turn off that certification. Yeah, let me know if you have any questions about it. should work anywhere because it's a browser browser tool. It's all just happening in your browser. Go check it out. Provide your feedback. We have many other tools there as well. We have, by the way, new tools coming. So, this one you can turn your microphone. You can turn your microphone on. Select the number of regions, brain regions. Yes, start microphone allow. And it will map the audio to your brain or the model of this brain. How cool is that? Yeah. So, yeah, you can check it out yourself. If it didn't work or something for you, do let us know. Do let us know. Can we have notebook as the acting CEO at the moment? Acting CEO of Barney Chaos. So, generate this audio for us. Play it while looking at the mind map. I put the mind map in full screen. Yeah, this is notebook. Should AI replace radiology readers? And it's a debate between two bots. And we are looking at this mind map while we Oh, it's 23 minutes. It's crazy. Let's do 25% um high uh playback speed. >> Welcome to the debate. Right now, um as we begin this conversation, a diagnostic radiologist somewhere in the world is sitting in a dark room. >> Yeah. Staring at a screen. >> Exactly. looking at a black and white image of a human lung and they have exactly three to four seconds to find a tumor before they have to, you know, clear the screen and move on to the next patient. >> 3 seconds. It's wild. >> It is 3 seconds. And because of that brutal, relentless math, human error rates in radiology haven't meaningfully improved in well 75 years, >> right? It's just stuck. >> Yeah. The daily clinical error rate is stubbornly frozen between 3 and 5%. So today we are asking a question that goes right to the heart of how we practice medicine. Is it time for artificial intelligence to fundamentally replace the human second reader in diagnostic radiology? >> And this whole discussion uh it emerges directly from the latest scientific literature. Right. >> Yes. Exactly. From clinical trial data and expert analyses on the implementation of deep learning in biomedical imaging. I represent the position that replacing the human second reader with AI is an entirely necessary collaborative paradigm. >> Okay. I mean it is the only way we can address systemic workforce shortages, profound human fatigue and the inherent hardwired perceptual errors of the human brain. >> Right? And I take the opposing view. While the speed and you know the efficiency games of artificial intelligence are incredibly alluring on paper, deploying these algorithms to autonomously replace human second readers is right now clinically unsafe. >> Unsafe. >> Absolutely unsafe. We are dealing with profound diagnostic accuracy risks, deeply embedded algorithmic biases, and a dangerous lack of generalizability when these tools uh leave the pristine conditions of a laboratory and hit the messy reality of actual hospitals. >> Well, let's unpack the core of why I believe replacing that second reader is absolutely necessary. Because if you're listening to this and wondering why brilliant, highly trained doctors are missing things, we have to talk about neurobiology. >> Sure, >> human readers are physically constrained. A radiologist suffers from visual fatigue, decision fatigue, and most importantly, something called inintentional blindness. >> Right. The brain filtering out what it doesn't expect to see. >> Exactly. When the human brain is intensely focused on a highly specific diagnostic task, say looking for a tiny pulmonary nodule, it actively biologically suppresses unexpected information in the peripheral vision, >> which is just a biological limit. >> Yeah, this isn't a lack of medical skill, you know, it's a hardwired limitation of how our optic nerve and cognitive processing load function. But when we implement AI in a hybrid workflow, specifically utilizing it to handle high volume, low ambiguity tasks, we see remarkable results. >> Remarkable in what way though? >> Well, look at the specific evidence from recent perspective trials in South Korea and European screening settings too when AI based computer aided detection or AICAD acts as the second reader in mimography. It has increased breast cancer detection rates by up to 13.8%. >> Wow. >> Yeah. 13.8%. And it achieved this without increasing the recall rate, which is critical. It didn't just flag every shadow and cause thousands of women unnecessary anxiety. It found actual cancers the human eye missed. >> Okay, but >> so the goal isn't replacing the radiologist entirely. It is replacing the second reader to create a tireless, highly sensitive partner. >> Look, I don't disagree that human error is a very real, very stubborn problem. I mean, 3 seconds in image is an absurd expectation for a human being. But substituting human error for machine error is not a viable solution because AI brings a completely different and utterly unpredictable set of failures to the table. >> Unpredictable how? >> Well, when a human doctor makes a mistake, we generally understand the cognitive mechanism behind it like the fatigue or the phobial vision limits you just described. We know how to correct for it. But when an AI makes a mistake, the why can be completely baffling. This is because of a massive vulnerability called shortcut learning. >> Ah, where the AI essentially cheats the test. >> Precisely. Deep learning models are at their core just math optimization machines. During their training, they look for the path of least resistance to lower their error rate. >> Right. They just want the right answer. >> Exactly. They don't understand biology. They understand pixel values. So they often find statistical correlations rather than actual causal biological mechanisms. >> Like what? For example, there are welldocumented cases where developers trained an AI model to detect a pumothorax, you know, a collapsed lung, but the AI actually just learned to detect the presence of a chest tube. >> Because the medical protocol for a collapsed lung is to insert a chest tube, >> right? The model wasn't detecting the incredibly subtle, wispy absence of lung markings that indicate a collapsed lung. It was detecting the bright, sharp, high contrast white pixels of the plastic tube inserted to fix it. >> It found the easy way out. >> Exactly. The math favors the sharp edges. We've seen similar issues where models learn to detect the radio marker of a portable X-ray machine to diagnose pneumonia. >> Wow. >> Why? Because patients who are sick enough to be in the ICU requiring a portable machine to be wheeled to their bed are statistically far more likely to have severe pneumonia. Because these models operate as opaque black boxes, replacing a human- second reader with an algorithm that lacks any contextual or causal understanding is simply irresponsible. >> I see why you think that, but let me give you a different perspective on how we evaluate diagnostic accuracy. >> Okay, go ahead. You're pointing out edge cases of shortcut learning which are real engineering challenges. I'll give you that. But if we pull back and look at the aggregate data, the performance is astounding. >> Is it though? >> It is. A massive meta analysis encompassing 86 different studies and over 129,000 medical images showed that AI achieved a pulled sensitivity of.92 and a specificity of.93. >> Right? >> To translate that into an insight, that data tells us AI isn't just matching human eyes. It's proving that human visual perception has a mathematical ceiling and we have finally built a ladder to climb past it. >> Okay, but >> and returning to the concept of inattentional blindness. AI never suffers from this. Think of the famous Harvard study where researchers embedded a tiny image of a gorilla into a stack of lung CT scans. >> Oh, I know this one, >> right? 83% of expert radiologists completely missed the gorilla. Why? Because they were instructed to look for lung nodules. Their brain filtered out the primate. >> Yeah, humans are hyperfocused. But the AI doesn't have a singular focus that blinds it to the periphery. It runs its mathematical weights across every single pixel with perfect tireless consistency every single time. >> I'm sorry, but I just don't buy that. Let me tell you why. >> Okay. >> The methodology behind those massive accuracy claims that 0.92 sensitivity is often fundamentally flawed. It comes down to how we define the ground truth in these studies. >> The ground truth. >> Yes. In many of the papers making up those massive meta analyses, the AI is simply benchmarked against another human radiologist's interpretation of a two-dimensional X-ray. Ah, I see. You're saying that AI is just being trained to successfully mimic a human's guess. >> Exactly. A 2D X-ray is just a shadow of a 3D object. The human is making an educated guess based on that shadow. >> Right. >> If a human missed a subtle underlying abnormality, and the AI also misses it, the AI is scored as correct in the study because it matched the human ground truth. It matched the flaw. >> Interesting. >> But when we benchmark AI against a true definitive medical gold standard, like a three-dimensional CT scan, the performance narrative completely falls apart. >> How so? Consider the recent prospect of study on wrist fracture detection. When measured against the definitive 3D CT scan, the AI generated 15 false positives and 13 false negatives. >> Okay. >> The human radiologist, only six false positives and six false negatives. When you force the AI to detect the actual underlying biology rather than just mimicking a human's interpretation of a 2D shadow, its performance drops significantly. >> That's an interesting point, though. I would frame it differently. >> How else can you frame a false positive rate that's more than double? Well, yes, the AI in that specific fracture study produced more false positives when compared to a 3D scan. But are human second readers perfectly accurate? Of course not >> fair. >> And more importantly, the discrepancy in false positives is completely manageable if we stop thinking of AI as a magic eightball that gives a definitive yes or no. We need to shift our conceptual framework to view AI as an uncertainty system. >> Explain what you mean by uncertainty aware in a clinical context. Think of the AI not as an absolute medical oracle, but as a geer counter for ambiguity. A modern, well-designed AI outputs a probability score, a confidence level. >> Okay? >> It sweeps over the obvious healthy tissue in silence. But the moment it detects a pixel pattern, it lacks mathematical confidence in, it clicks. It flags the human. >> So, it just points things out, >> right? It doesn't formally diagnose the anomaly. It simply maps the boundaries of its own ignorance. If it flags a low confidence read, which might just be a false positive, it escalates that specific image to the human radiologist. >> I see. >> In that hybrid workflow, a false positive doesn't lead to an unnecessary surgery. It simply acts as a highly sensitive safety net that forces the human to look closer. It optimizes the human 3 seconds by focusing their eyes exactly where the ambiguity lies. But that entire workflow assumes the AI's confidence score is reliably calibrated in the first place. And that brings us to the root cause of these accuracy failures, which is the underlying training data. >> You mean the bias? >> Yes. The reason an AI might generate wild false positives or confidently miss a fracture without triggering your metaphorical geer counter is directly tied to algorithmic bias and data limitations. We have a severe issue with geographical data access. >> Sure, regional differences, >> right? An AI model trained exclusively on patient data, specific scanner protocols, and demographics from a massive academic hospital in Boston might look incredibly accurate on paper, but when you deploy that exact same algorithm to a community clinic in Cleveland or a diverse patient population in San Francisco, it can fail catastrophically. >> This is known in the field as the out of distribution problem. >> Precisely. The AI overfits to the spirious artifacts of its training environment. It learned what a healthy lung looks like on a specific General Electric scanner in Boston, not what a healthy lung looks like objectively. >> Right. And my deep concern here is regulatory oversight. We are seeing bodies approving hundreds of these algorithms, often classifying them as software as a medical device, perhaps far too quickly. >> They're moving fast. >> They are clearing these tools before we have resolved how brittle they are to non-cultural and regional differences. Deploying them as autonomous second readers when they fail the moment you move them across state lines is incredibly premature. I will readily acknowledge the challenge of data scarcity. Developing robust models requires massive diverse data sets. And as a medical community, we are rightly constrained by stringent patient privacy regulations. >> Yeah, HIPPA and GDPR. >> Exactly. You can't just email a million patient X-rays across the world. Aggregating that data is incredibly difficult. But the engineering community is already solving this. >> Solving it how? >> The solution to the data bottleneck is synthetic data. >> Generating fake medical records to train real medical algorithms. Not just fake records, mathematically precise synthetic biomedical data. For listeners who might not be deep into machine learning, developers use something called generative adversarial networks or GANs. >> The cat and mouse game, >> right? Think of it as two AI networks playing a game. One network acts as a master forger trying to create a completely artificial X-ray from scratch. The other network acts as the detective trying to spot the forgery. >> Okay. >> Over millions of cycles, the forgeries become mathematically indistinguishable from reality. We're seeing platforms like Bioeneagos generate synthetic EEG waveforms and complex periodic noise. >> But does it work for imaging? >> Yes. In imaging, this allows us to train algorithms on vast, perfectly annotated data sets that represent every conceivable pathology variation without ever compromising a real patient's privacy. We can even intentionally inject artificial noise into the training data to simulate the exact scanner variations from Cleveland or San Francisco that you were just worried about. >> I come at it from a different way. I have to strongly question the true clinical usefulness of synthetic data. >> Why? It's perfectly annotated. >> But can a mathematically generated synthetic image truly replicate the chaotic, messy, unpredictable artifacts of a real human body? Biology is infinitely complex. >> It is, but the metals capture that complexity. >> Do they? When we synthesize data, we are building it based entirely on our current limited human understanding of a disease. Therefore, training an AI on synthetic data runs the massive risk of simply baking our own human blind spots directly into the algorithm's foundation. >> I think that's a bit pessimistic. >> Not really. We aren't teaching the AI to discover new biological truths. We are trapping it inside an echo chamber of our own assumptions. It creates a false sense of security. The AI looks incredibly robust in contesting right up until it encounters a real human being with a novel presentation that the synthetic generator never even imagined. >> But you have to view this through the lens of actual clinical deployment. It's easy to demand perfection when we're talking about massively resourced academic centers with armies of specialists, >> right? >> But contrast that with smaller rural critical practices, the rural mom and pop shops. You are arguing against AI as a second reader by highlighting its imperfections. But in many underserved clinics, they don't have the staffing budget to even have a human second reader. >> I understand the shortage. >> For them, AI isn't replacing a human colleague. It is providing one. >> That assumes the rural clinic has the capability to safely interact with the AI >> and they increasingly do. We are seeing cloud connectivity and satellite internet like Starlink being deployed in remote areas even in places as underserved as RO clinics in Uganda. >> Okay. Starlink helps with internet. Sure. >> Right. And this technology allows a sole practitioner working in total isolation to instantly access top tier AI diagnostic support for that doctor. An uncertainty AI that flags a suspicious shadow on a chest X-ray is an absolute gamecher. It democratizes expert level analysis across the globe. That's a compelling argument, but have you considered the hidden infrastructure divide that makes deploying this technology so dangerous in those exact settings? >> Infrastructure divide? You mean the internet? >> No, I mean the IT plumbing inside the hospital. Big academic centers have sophisticated IT infrastructure. They utilize advanced data interoperability standards like DICOM segmentation objects and HL7 protocols to feed data into TED 1500 structured reports. >> Wait, let me stop you there. I understand what a heat map or an overlay is. If an AI finds a tumor, it highlights it. But I'm a bit confused. Why can't a rural clinic just look at the same heat map that a Boston hospital uses? What is actually breaking down in the plumbing there? >> It's a great question and it comes down to how the software interacts with the image. Think of DICCom SEG like a layered Photoshop file. >> Okay. >> When the AI flags a tumor at a major university hospital, it creates a fractional segmentation, a probabilistic heat map that exists on a separate interactive layer over the original X-ray. >> So the doctor can manipulate it. >> Exactly. The radiologist can toggle it on and off, adjust the opacity, query the data, and critically evaluate what's underneath it. But a small rural clinic usually doesn't have that expensive pack infrastructure. They just receive a flattened standard secondary capture image >> like a JPEG. >> Exactly like a JPEG. The AI's colored box is literally burned into the image pixels. The doctor can't turn the overlay off. >> They can't interrogate the data to see the underlying tissue the AI is obscuring. And this forces the overworked soul practitioner into a dangerous psychological trap known as automation bias >> where they just cognitively offload the decision to the machine >> precisely. They are exhausted, they are alone, and the machine has drawn a definitive unreovable box on the screen. A landmark study out of Harvard looking at 140 radiologists proved exactly how dangerous this is. >> What did they find? >> They found that while a highly accurate AI tool boosted human performance, a poorly performing AI tool actually actively worsened the accuracy of the human clinicians using it. >> Wow. So it made them worse. >> Yes. The human doctors stopped trusting their own eyes. They accepted the machine's false positives and ignored their own instincts. If you deploy a biased AI to a rural clinic that lacks the IT infrastructure to critically evaluate the data layers, you aren't providing a helpful colleague. You are providing a massive clinical liability. >> I'm not convinced by that line of reasoning because it assumes we will deploy AI recklessly without adapting our workflows. Let me offer a different analogy. Think about a smart spell checker on a word processor. >> Okay, a spell checker. >> We don't blindly accept every grammar suggestion it makes. The software handles the routine checking, the high volume, low ambiguity tasks, but it only highlights the words it mathematically doesn't recognize. >> Right? >> The human writer still has to look at the red squiggly line and make the final contextual decision. The future collaborative paradigm in radiology works the exact same way. The AI operates as an intelligent triage system. >> A triage system. >> Exactly. If that rural physician in Uganda has a queue of 50 chest X-rays, the AI can confidently clear the 40 definitively healthy ones and escalate the 10 complex ambiguous cases directly to the top of the doctor's screen. We are safely optimizing their severely limited time by letting the machine do the heavy lifting, keeping the human firmly in the loop for the complex clinical reasoning. >> But a spell checker is dealing with the rigid universal rules of human grammar, and AI and radiology is dealing with the infinite variability of human biology. >> It's still anomaly detection. Yes, but if the AI confidently clears a chest X-ray because it falls victim to shortcut learning, perhaps missing a tumor because the patient's shoulder blade cast an unusual shadow on a cheaper X-ray machine, that patient goes home with an undiagnosed cancer. A spellch checker error is a typo. A radiology error is a tragedy. Which is exactly why the continuous engineering of that uncertainty awareness is the key. If we look at the core of this debate, the reality is that the human neurobbiological capacity to interpret thousands of black and white pixels at a rate of one every 3 seconds has reached its absolute limit. I agree with the limit. >> Right? So AI is fundamentally necessary to move us past that 75-year plateau. By replacing the traditional human second reader with a highly sensitive uncertainty AI system, we can democratize expert level analysis across all clinical settings from massive academic hospitals to remote rural clinics. we can catch the subtle pathologies that human eyes simply filter out all while giving doctors their time back. >> And to summarize my position, while the efficiency gains and the potential for democratizing care are undeniably alluring, the profound risks remain entirely unresolved. Unresolved for now, maybe the dangers of shortcut learning, where algorithms latch on to speirious correlations like chest tubes rather than biological truths, are real. The geographical data bias that causes models to fail when deployed in new demographics is a massive vulnerability. Right? And our reliance on weak human-drived ground truth validation creates a false sense of security. Until we solve the core problem of how these machines actually know what they are looking at, AI cannot be blindly trusted as an autonomous second reader with rigorous continuous human skepticism. >> Well, I think where we absolutely converge today is on the future role of the radiologists themselves. The old anxiety maybe 5 years ago was that AI would replace human radiologists entirely. >> Yeah. The robot doctor panic. >> Exactly. But the consensus today is quite different. AI won't replace radiologists, but radiologists who effectively use AI will undoubtedly replace those who do not. >> I agree completely. The future isn't a battle of human versus machine. It's about defining the safest, most rigorous boundaries for human machine collaboration >> lens and a clinical safety lens for our listeners who are deeply engaged with a technical implementation side. There is so much more to explore in the source material regarding the vital role of data standards. >> Yeah, like diccom seg, >> right? Understanding how DICOM, SEG, and HL7 actually function to make this interoperability safe across different hospital systems. >> The data architecture matters just as much as the algorithm itself. >> Absolutely. We will leave it to you, the listener, to form your own conclusion on where the balance of risk and reward truly lies. But as you do, think about what we truly want for medical imaging. We want absolute visible certainty >> always. But in the murky, infinitely complex reality of interpreting human biology, perhaps the greatest tool we can build isn't one that claims to know everything. Perhaps the greatest tool we can build is one that finally knows exactly what it doesn't know. >> There's a human in the loop. Just if you were wondering, there is a human in the loop. Hey, check out bcaos.com. Provide your feedback. I'll see you next time. Bye.