Deep Learning-Based Sex Classification and Chronological Age-Group Estimation from Knee Radiographs Using a Multi-Task Framework with Rank-Fusion and Tri-Loss Learning

In the world of forensic science and medical diagnostics, identifying an unknown person often begins with a simple question: who were they? Two of the most fundamental answers are their biological sex and their age. For decades, experts have relied on the skeleton to provide these clues, examining bones for signs of how they were built and how they have changed over time. The knee joint, a complex hinge where the thigh bone meets the shin bone, is particularly valuable for this work. It is a site that shows clear differences between men and women, and it undergoes predictable, gradual changes as a person moves from young adulthood into old age. Traditionally, doctors and forensic specialists have looked at X-rays of this joint with their own eyes, using standardized charts to score the bone's condition. However, this manual process can be slow, and different experts might not always agree on the same image. As the field of artificial intelligence has grown, researchers have begun to ask if computers can learn to see these subtle patterns more consistently and accurately than the human eye.

A team of researchers from Kütahya Health Sciences University and Kütahya City Hospital in Turkey has taken a significant step forward in this effort. They developed a new computer system designed to read knee X-rays and automatically determine both the sex and the age group of the person who owns that knee. Instead of relying on a single method, they built a framework that handles two distinct tasks at once. For the first task, figuring out if the knee belongs to a man or a woman, the team did not trust just one computer model. Instead, they trained three different types of advanced artificial intelligence systems to look at the same image. They then combined the results of these three systems into a single, unified decision. This approach, which they call a Rank-Fusion Ensemble, allowed the computer to weigh the strengths of each model, resulting in a highly reliable identification of biological sex.

The second task was more complex: estimating the person's age. Unlike sex, which is a binary category, age is a continuous progression. A person in their thirties shares many physical traits with someone in their forties, making the boundaries between age groups blurry. To handle this, the researchers used a specialized computer architecture called ConvNeXt-Tiny, which is particularly good at spotting fine details in images. They taught this system to understand that age groups follow a specific order. They designed a training method that penalized the computer less for guessing a neighbor's age group and more for guessing a distant one. For example, if the system guessed a fifty-year-old was forty, that was considered a minor error; guessing they were twenty would be a major mistake. This approach, which the authors call a Tri-Loss strategy, helped the model respect the natural, step-by-step nature of aging.

The team tested their system on a carefully selected collection of 1,042 knee X-rays taken from patients between the ages of 22 and 69. The data was balanced so that there were nearly equal numbers of men and women, and the patients were spread evenly across five different decades of life. When the system was tested on images it had never seen before, the results were striking. For determining sex, the combined system was correct 96.65% of the time. It was exceptionally good at distinguishing between male and female bone structures, achieving a score that indicates near-perfect reliability in separating the two groups. This performance was significantly better than any of the three individual computer models used to create the ensemble.

The age estimation task proved more challenging, as the differences between adjacent decades are often subtle. Even so, the specialized system achieved an accuracy of 71.29% in guessing the correct decade. More importantly, the system rarely made wild guesses. When it was wrong, it was almost always off by just one decade. The researchers found that in 96.65% of all cases, the computer's guess was either the correct age group or the one immediately before or after it. It never made the kind of extreme error where it would mistake a young adult for an elderly person. This suggests that the system has successfully learned the gradual, overlapping changes that happen in the knee as a person ages, rather than just memorizing distinct categories.

The study also looked at how confident the computer was in its answers. It found that when the system expressed high confidence, it was usually right, meaning the probabilities it calculated were trustworthy. This is a crucial detail for medical and forensic use, where knowing the reliability of a result is just as important as the result itself. The researchers noted that while the system performed exceptionally well at identifying the very young and the very old, the middle age groups were slightly harder to distinguish, likely because the physical changes in the knee during those years are more gradual and less distinct.

This work demonstrates that deep learning can serve as a powerful, objective tool for reading the story written in our bones. By combining multiple artificial intelligence models for sex identification and using a specialized approach for age estimation, the researchers have created a system that offers a quantitative alternative to traditional, manual observation. While the study focused on a specific population and a specific range of ages, the results suggest that automated analysis of knee radiographs can provide consistent, high-accuracy demographic profiles. This capability could eventually support forensic experts in identifying unknown remains or help clinicians assess age-related changes in living patients, all without the variability that comes from human interpretation. The findings confirm that the knee, with its clear sexual dimorphism and predictable aging process, remains a rich source of information, now accessible through the lens of modern machine learning.