Trustworthy Few-Shot Transfer Learning withExplainable AI for PCOS Ultrasound Classification

In the quiet, dimly lit rooms of a modern clinic, a doctor holds a probe against a patient's skin, watching a grainy, black-and-white screen come to life. This is an ultrasound, a window into the body that reveals the shape and texture of internal organs. For women of reproductive age, one of the most common conditions doctors look for is polycystic ovary syndrome, a hormonal disorder that can affect fertility and overall health. To diagnose it, a specialist must count tiny fluid-filled sacs, called follicles, on the ovaries and measure their size. This task is difficult. It relies heavily on the doctor's experience, the quality of the machine, and even the angle of the probe. Two experts looking at the same image might disagree on the count, leading to uncertainty in treatment.

For years, scientists have tried to build computer programs that can read these images with the same skill as a human expert. These programs, known as artificial intelligence, are usually trained by showing them thousands of labeled examples until they learn to recognize patterns. However, in the real world of medicine, gathering thousands of perfect, labeled images is often impossible. Hospitals may have only a few dozen cases, or the data might be scattered across different clinics. This creates a major hurdle: how can a computer learn to diagnose a condition when it has very little to study? Furthermore, even when a computer makes a correct guess, doctors need to know why. They need to see which part of the image the computer is looking at, to ensure it is not just guessing based on a random speck of noise. This need for clarity is the heart of a new study that explores how to build trustworthy medical AI when data is scarce.

A team of researchers from Daffodil International University in Bangladesh set out to solve this double problem. They wanted to know if a computer could learn to identify polycystic ovary syndrome from ultrasound images using only a handful of examples, and whether the reasons it gave for its decisions would remain reliable under those difficult conditions. To do this, they used a massive collection of 11,784 ultrasound images, split into two groups: those showing the syndrome and those that did not. Instead of training their models on the entire collection at once, they simulated a scenario where the computer had to learn from very small groups of images. They tested the system with as few as five examples per condition, then gradually increased the number to ten, twenty, fifty, and finally the full dataset.

The researchers compared two different types of computer architectures, which can be thought of as the underlying engines that process the visual information. One engine was designed to be extremely efficient but was kept mostly frozen, meaning it could not change its internal structure much during learning. The other engine was allowed to adjust its final layers, giving it more flexibility to adapt to the specific details of the medical images. The results showed a clear winner in the low-data environment. The flexible engine, which adjusted its deeper layers, consistently outperformed the rigid one. When given only five examples to learn from, the flexible system correctly identified the condition about 79 percent of the time, while the rigid system managed about 76 percent. As the amount of training data grew, both systems improved, but the flexible one maintained a steady lead, especially when the number of examples was small. At the lowest data levels, the rigid system achieved a respectable performance level of 75.5% accuracy and 0.844 AUC, though it showed slightly lower accuracy and a less balanced ability to distinguish between the two types of images compared to the flexible system.

Perhaps the most surprising discovery concerned the "why" behind the computer's decisions. The researchers used a technique that highlights the specific areas of an image that influenced the computer's choice, creating a heat map that glows over the relevant parts. They wanted to see if these heat maps became blurry or unreliable when the computer was trained on very few images. They measured the trustworthiness of these explanations using several rigorous tests, checking if the computer's confidence dropped when the highlighted area was removed, and how stable the heat map remained when the image was slightly altered. The findings were reassuring. While the specific metrics varied between models and data amounts, the overall quality of the explanation did not statistically degrade as the amount of training data shrank. Even when the computer was trained on just five examples, the heat maps it produced remained statistically faithful in their fidelity, showing that the computer was genuinely looking at the follicles and the ovarian structures that matter to a doctor, regardless of how little data it had seen, even if the visual activation patterns appeared somewhat more diffuse than in full-data scenarios.

This stability held true even when the researchers tested the system's robustness against noise, simulating the kind of variations that happen in real-world clinics where equipment differs from hospital to hospital. The flexible system remained steady, while the rigid system showed signs of sensitivity, its explanations shifting more easily when the image quality changed. The study suggests that for medical AI to be truly useful in resource-limited settings, where large datasets are unavailable, the key is not just to have a powerful model, but to choose a model that can adapt its deeper layers to the specific task. The research confirms that it is possible to build systems that are both accurate and explainable, even when the training data is extremely limited. This offers a path forward for deploying trustworthy diagnostic tools in places where every single patient image counts, ensuring that the computer's reasoning remains clear and reliable for the doctors who depend on it.