Automated Forensic Sex Determination from Skull Morphology Using CT and AI

By Aryaman PattnayakReviewed by Susha Cheriyedath, M.Sc.Dec 3 2023

A study published in the journal Scientific Reports demonstrates how an artificial intelligence technique leveraging Computed Tomography (CT) image analysis of deceased individuals' skulls can reliably determine sex based on subtle bone shape differences. The automated approach promises greater objectivity and reproducibility compared to inconsistent visual examinations by forensic examiners that require extensive specialized experience.

Attempts to identify unknown skeletal remains often depend critically on assessing sexual characteristics from minute variations in skull morphology and bone shapes. Through detailed anatomical familiarity built over years of practical casework, forensic anthropologists train their perceptual skills to discern reliable signals within complex osseous structures indicating the deceased's biological sex.

However, such subjective evaluations by specialists probing photographs or directly palpating bones frequently suffer from irreproducibility issues across practitioners in the same case. Disagreements between examiners' sex judgments are not uncommon, even for seasoned experts. These reliability gaps become further pronounced for novice forensic trainees needing more proficiency at teasing out the nuanced clues within intricate and sometimes damaged bone geometries.

Since criminal investigations hinge on accurate biological profiles, qualitative methods prone to subjective errors pose a severe impediment. Integrating quantitative computerized solutions that ensure standardized objectivity is therefore essential. Recent studies have shown promise in testing automated machine learning algorithms capable of parsing volumetric CT data to classify sex. However, most develop and validate approaches using convenience datasets of living subjects rather than real deceased individuals undergoing analysis.

Applicability to actual corpses thus remains unproven. Practical forensic casework regularly contends with extensive tissue decay from longer postmortem intervals. Generated gases and fluids can warp shapes that machines learn to depend on—ignoring such reality gaps during training risks models to underperforming when deployed on target beneficiaries. So, research centered on cadaver CT data itself is critical.

About the Study

Striving to develop a pragmatic forensic assistant, a group of scientists from Kyoto Prefectural University of Medicine in Japan implemented an advanced deep neural network leveraging attention mechanisms to categorize deceased individuals as male or female based solely on intrinsic three-dimensional skull structure cues extracted from accurate full postmortem CT scan volumes.

Through a multi-stage process, the researchers first compiled an expansive dataset from over 2000 East Asian cadavers' CT series archived at their institution, split evenly by sex. Extensive preprocessing digitally isolated and normalized skull anatomy from the surrounding cervical vertebrae and residual brain matter obscuring surfaces to highlight critical regions.

Following dataset partitioning for training and Performance testing, the team tapped recent breakthroughs in computer vision to configure a specialized artificial intelligence architecture for analyzing shapes. A dense connectivity convolutional neural network handled automatic feature extraction from skull patches to highlight differences. Subsequent attention layers assigned scalar importance weights to each patch according to sex discrimination ability.

The researchers finally evaluated the trained model's performance in labeling reserved testing cases against assessments by three practicing forensic examiners with varying skill levels. The algorithm's predictions were measured against ground truth sex to ascertain classification accuracy. Its attention visualizations also offered clues into morphological focus areas.

Experimental Results

After optimizing training hyperparameters, the machine learning approach achieved an overall testing accuracy of 93% in predicting whether a given unknown skull was male or female. This performance significantly exceeded the ~83% and 77% testing accuracy attained by two seasoned specialists with years of casework proficiency. A beginner trainee's 63% correctness lagged even further behind the automated technique.

The model proved equally proficient for both sexes, demonstrating 92% and 95% female sensitivity and specificity. The consistently high metric across categories signals reliable application without inherent gender biases that handicap human evaluations. Meanwhile, attention mapping primarily highlighted the mandible as notable for predictive decisions. This aligns with anthropological understandings of strong sexual dimorphism in the jaw.

Broader Implications

Together, these promising results firmly establish the viability of the proposed machine learning paradigm as a forensic assistant amplifying identification accuracy alongside traditional morphoscopic analysis - even for specialists with extensive expertise.

By extensively pre-training on actual postmortem cases exhibiting chewing artifacts and bacterial decay rather than convenient clinical scans, the practical method further avoids limitations plaguing prior work reliant on living anatomy. The techniques' independence from subjectivity or experience levels marks a sea change from established but flawed protocols necessitating years of case exposure to attain competence.

Future iterative models trained on exponentially more extensive and more diverse ethnic populations and traumatic injury patterns should only further expand usefulness across unidentified remaining challenges. However, the seminal study already conclusively demonstrates the immense feasibility and upside of forensic anthropology integrating sophisticated AI to modernize subjective sex evaluation tasks that confront reproducibility obstacles.

Innovative Deep Learning Approach

The research pushes boundaries on multiple fronts to realize a solution well beyond convention. The inventive configuration combining CT input spaces with cutting-edge deep neural manipulation uniquely suits the problem. Volumetric density scans capture sex markers eluding static photographs. Dense connectivity networks handle high-dimensional datasets. Attention weighting accentuates influential features for decisions. Moreover, multiple-instance learning efficiently trains complex models on limited specialized data.

Together, these advances facilitate nuanced parsing of intricate morphology. Each complementing innovation provides incremental advantages that compound substantial forecasting improvements over prior attempts with other machine learning classifiers on related tasks. The future integration of even more advanced architectures as they continue to emerge across computer vision spheres will likely only continue yielding better interfaces at the intersection of osteology and technology.

Practical Functionality

At its core, the research produces a promising proof-of-concept for AI-expanding forensic anthropological evaluations facing reliability gaps for urgent identifications. Rather than outright replacing trained specialists, the proposed assistance paradigm strictly bolsters the probabilities of accurate casework alongside standard workflow.

The tool outputs independent sex predictions from non-subjective analysis of morphological features before returning attention maps conveying contributing areas to the human expert, who still retains ultimate investigative authority. This transparent approach incorporates computational forensic insights during case reporting without entirely ceding conventional processes. Immediate applications could scan unidentified remains that examiners subsequently analyze, informed by, or resolve any discrepancies from the algorithm interpretation.

Long-term integration could even readily operate on-site if deployed onto mobile stations receiving CT feeds for rapid augmented evaluation, facilitating triage at mass disasters. Either adoption path would synergize respective strengths - the innate but inconsistent pattern recognition of seasoned practitioners and the unbiased analytical computations of algorithms. Further technological progress will only continue to shuffle this complementary balance towards increasingly centralized automation.

Conclusion

This breakthrough research provides an initial framework seamlessly integrating state-of-the-art deep learning alongside CT imaging to establish a readily automated and objective forensic sex assessment tool delivering performance markedly surpassing humans. The technique promises to amplify identification accuracy over inconsistent visual protocols requiring years of specialized case exposure to attain competence. Broader development of such versatile computational forensic aids leveraging leading AI inventions will prove indispensable assets streamlining and standardizing challenging casework still reliant on antiquated morphotropic evaluations.