Cross-center sex estimation from patellar CT: multi-feature transformer fusion of morphometry, radiomics, and deep features
This study presents a robust multi-feature Transformer fusion framework that integrates patellar morphometry, radiomics, and deep learning features from CT scans to achieve highly accurate and generalizable sex estimation, outperforming single-feature models and forensic experts while significantly reducing processing time.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the quiet aftermath of a disaster or a crime, forensic scientists face a daunting puzzle: who was this person? The first and most critical step in solving that puzzle is determining the sex of the remains. For decades, experts have looked to the pelvis and the skull, the bones that most clearly show the physical differences between males and females. But in the chaotic reality of mass disasters or ancient sites, these bones are often broken, burned, or missing entirely. When the usual clues vanish, investigators must turn to other parts of the skeleton. One such survivor is the kneecap, a small, dense bone that often withstands fire and trauma better than the rest of the body. While it is smaller and less obvious than the pelvis, it carries a hidden map of biological sex within its shape and texture, waiting to be read.
A team of researchers from Qinghai University and affiliated hospitals in China has developed a new way to read that map. They created a computer system that acts like a super-observer, combining three different ways of looking at a kneecap to guess whether it belonged to a man or a woman. Instead of relying on just one method, the system merges simple measurements of size, a detailed analysis of the bone's internal texture, and a deep, abstract understanding of the bone's overall form. By teaching a computer to weigh all these clues together, the researchers found a way to make sex estimation faster, more consistent, and more accurate than current human methods, even when the bone scans come from different hospitals with different equipment.
The study began with a large collection of knee scans from two hospitals in Qinghai Province. The researchers gathered images from 836 people, carefully selecting those with healthy knees and clear images. They split these into two groups: a larger group to teach the computer how to recognize patterns, and a smaller, separate group to test if the computer could apply what it learned to new, unseen data. To build their tool, the team first manually outlined the kneecap on each scan, isolating it from the rest of the knee joint. From this isolated bone, they extracted three distinct types of information. First, they measured the bone's basic geometry, calculating its length, width, thickness, volume, surface area, and the distance around its edge. Second, they used a technique called radiomics to break the image down into hundreds of tiny texture details, capturing the grain and pattern of the bone's interior that the human eye cannot see. Third, they used advanced artificial intelligence models to scan the bone and extract deep, abstract features that represent the bone's overall shape and structure in a way that mimics how a human brain might recognize a face or an object.
The core of their innovation was how they combined these three streams of information. Rather than simply averaging the results or stacking the data on top of each other, they used a sophisticated computer architecture known as a Transformer. This system works like a dynamic decision-maker that learns to pay attention to the most important clues for each specific bone. It can decide that for one kneecap, the texture is the most telling sign, while for another, the overall size matters more. This flexibility allows the system to adapt to variations in the data, such as differences in how the scans were taken. The researchers tested this system against simpler models that used only one type of information, as well as against other ways of combining the data.
The results showed that the combined approach was superior. On the internal test set, the system correctly identified the sex in 94.3% of cases, a significant improvement over the single-method approaches. More importantly, when tested on the external group of patients from a different hospital with slightly different scanning equipment, the system maintained a high level of accuracy, correctly identifying the sex in 87.9% of cases. In contrast, the simpler models that relied on just one type of data saw their accuracy drop more sharply when faced with these new conditions. The study also compared the computer's performance to that of human experts. Three human readers, ranging from a student to a board-certified forensic specialist, manually measured the same 15 bones using traditional formulas. While the human experts were generally accurate, their measurements varied significantly from one another; for the same bone, one expert might measure the length as 40 millimeters while another measured it as 48 millimeters. The computer, however, produced the exact same result every time, eliminating this human inconsistency. Furthermore, the computer completed its analysis in an average of 30 seconds per bone, whereas the human experts took anywhere from four to thirteen minutes each.
The researchers noted that the thickness of the scan slices played a crucial role in the performance of different methods. When the scans were very thin, capturing fine detail, the texture-based and deep-learning methods performed exceptionally well. However, when the scans were slightly thicker, as is common in some real-world scenarios, the performance of those complex methods dropped. The simple size measurements remained stable regardless of the scan thickness, which is why the system's ability to blend all three methods was so effective; it could rely on the stable measurements when the detailed texture data became less reliable. The study concludes that this multi-feature approach offers a robust, objective tool for forensic identification. It suggests that by letting a computer learn to weigh different types of evidence, investigators can get a more reliable answer about a person's sex, even when the bones are damaged or the data comes from varied sources. While the study was limited to a specific population in China and used scans from living people, the authors believe this method provides a strong foundation for future work that could help identify victims in mass disasters or analyze remains in archaeological contexts, turning a small, often overlooked bone into a powerful key for identification.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.