Cancer: How AI Really Helps the Radiologist

On a mammogram, a tumor invisible to the naked eye hides within the tissue. An algorithm spots it before it even becomes palpable. It all begins between 2015 and 2020. The first algorithms learn to read medical images and to flag suspicious anomalies within them. The COVID pandemic speeds everything up, between 2020 and 2025. The FDA clears more and more medical devices built on artificial intelligence. A medical image is thousands of shades of gray. A radiologist's eye grows tired; the machine never wearies. Before any analysis, the image is normalized. It is brought to the same scale and the same contrast, so we compare like with like. At the heart of the system, convolutional neural networks. On their own, they extract the patterns of an image, without our having to describe them. These networks excel on high-resolution images, like a CT scan or an MRI. They catch patterns a human might miss. More recently, another architecture emerges: the Vision Transformers. Where the CNN looks piece by piece, they analyze the whole image at once. Their secret: an attention mechanism. The model links each region to the others, and finally grasps the overall context of the entire image. Breast cancer benefits first. Google Health's model and Lunit INSIGHT MMG analyze mammograms in search of tumors. On well-defined tasks, these systems match a single radiologist, sometimes surpass them. But it all depends on the exact evaluation protocol chosen. The consequence is concrete: more cancers detected early. And fewer needless biopsies, when the machine clears a doubt that turns out benign. Lung cancer, so hard to see early, moves forward too. Aidoc and AI-Rad Companion, from Siemens Healthineers, scan chest CT images. They track down tiny lesions, the sign of an emerging cancer. By spotting them, they reduce false negatives, those cancers that go unseen. The same method goes well beyond cancer alone. Today it is applied to diabetic retinopathy, to bone fractures, and to heart conditions. Here is how the system actually works, step by step. The model never rules: it proposes suspicious areas on the image. To each area, it assigns a confidence score. A rating that says: here is how sure I am of being right. Then it is the human radiologist who reviews these areas. They confirm the real alerts themselves, and set aside those that are not. A false positive set aside is never wasted. It goes back into training, and the model learns not to repeat the mistake. And this is what matters most: the final decision stays medical. The machine sharpens the radiologist's eye, it does not replace it. Behind these tools, a whole ecosystem is at work. The giants Google Health, IBM Watson Health, and Microsoft invest heavily in these new models. Alongside them, highly specialized startups advance. Aidoc, Zebra Medical, and Lunit move fast on very precise diagnostic niches. And the imaging manufacturers complete the picture. Siemens Healthineers, GE Healthcare, and Philips build this software directly into the heart of their machines. But none of this is truly without risk. A model faithfully reproduces the biases of the data it was trained on. If the training data lacks diversity, then the diagnosis goes off course. The underrepresented populations become the first victims of that error. Faced with these stakes, regulation moves forward. The FDA imposes strict frameworks to assess each device before it reaches the market. These certification processes are long and complex to clear. They protect the patient, but they can also slow the arrival of useful innovations. One last obstacle remains, more human than technical. The radiologist must be convinced the tool is reliable, to adopt it day to day. What has changed is not that the machine decides. It is that it now sees what the eye alone could no longer see. The real revolution is here: a radiologist augmented, not replaced. The human and the model finally look at the same image, side by side. At Koaee, we follow these advances very closely. They remind us of one requirement: a useful AI assists the human, without ever deciding in their place. It is with this same rigor that we design our conversational assistants. Want to dig deeper? The full article awaits you in the comments. Are you passionate about augmented medicine, and artificial intelligence in general? Subscribe to this channel, like the video, and meet us at koaee.ai.